PREFACE This volume contains the proceedings of the International Conference on Harmonic Analysis and Related Topics held at Macquarie University, Sydney, from January 14–18, 2002. The conference celebrated the many significant achievements and contributions to mathematics of Professor Alan G. R. McIntosh, on the occasion of his sixtieth birthday. A large number of the speakers were his mathematical proteg´es and co-authors, and all expressed their gratitude to Alan for the way he had influenced or nurtured their careers. The conference brought together researchers working in harmonic and functional analysis, linear and non-linear partial differential equations, functional calculus and operator theory, and included researchers from North America, Europe and Asia as well as Australia. We gratefully acknowledge the support of the contributors to this volume and the sponsors of the conference, the Centre for Mathematics and its Applications at ANU and the Department of Mathematics and the Division of Information and Communication Sciences at Macquarie University. Our thanks also go to Dr Chris Wetherell for his editorial assistance in preparing this volume. We especially thank Alan McIntosh for his participation. A short synopsis of Alan’s career, and a full list of his publications to date are included in this volume. Contributions for the proceedings were invited from all participants. All papers received were carefully refereed by peer referees. Xuan Thinh Duong, Alan Pryde (Editors)
i
FOREWORD Alan McIntosh spent his undergraduate years at the University of New England in Armidale NSW where he obtained his BSc(Hons) in 1962. He gained his PhD at the University of California, Berkeley in 1966 under Frantiˇsek Wolf. After a year at the Institute for Advanced Study, Princeton, he returned to Australia and began a long association with Macquarie University. In 1999, after 32 years at Macquarie, he became Head of the Centre for Mathematics and its Applications at ANU. Throughout his time at Macquarie, McIntosh provided leadership in analysis. For many years his group ran weekly seminars which typically attracted participants from the other universities in Sydney. Many well-known mathematicians presented lectures. McIntosh made strong post-doctoral appointments with a diversity of backgrounds and fostered excellence in research. One of his major goals was and remains to nurture young mathematicians. The result has been a long list of outstanding mathematical advances from McIntosh and those who have been fortunate to come under his influence. McIntosh’s early work on accretive bilinear forms was heavily influenced by Tosio Kato (Berkeley), and it was a problem posed by his mentor in 1960 that led to McIntosh’s most significant work. The Kato square root problem asks whether the square root of an accretive operator in divergence form is stable under perturbations of the original operator. It was not until 1981 that the one dimensional version of this problem was solved in the fundamental work of Ronald Coifman (Yale), McIntosh and Yves Meyer (ENS-Cachan). In this paper [11], the authors also solved the conjecture of Calder´on on the boundedness of the Cauchy integral on a Lipschitz curve. It took until 2000 for the two dimensional version of the problem to fall at the hands of Steve Hofmann (Missouri) and McIntosh. In the following year, the full arbitrary dimensional solution was given in the joint work of Pascal Auscher (Paris-Sud), Hofmann, Michael Lacey (Georgia Tech.), McIntosh and Philippe Tchamitchian (Marseille). Beginning with his collaboration with Coifman and Meyer, McIntosh was to forge remarkable links between the harmonic analysis of the Zygmund school and the operator theory of Kato and others. This work involves the use of square function estimates associated with particular operators, and the construction of the corresponding functional calculi, and depends ii
iii
upon Lp estimates for singular integrals. One specific aim is to study boundary value problems for linear partial differential equations with non-smooth coefficients on irregular domains, and associated nonlinear problems. Results are obtained under natural geometric conditions and these are of special interest when applied to nonlinear problems arising from physical or geometric phenomena. They also have implications for parabolic and hyperbolic problems. The principal methods involve developing the harmonic analysis of operators directly on domains or on their boundaries. Topics related to McIntosh’s research include boundedness of singular integrals and Fourier multipliers on Lipschitz surfaces; heat kernel bounds and functional calculi of elliptic partial differential operators; compensated compactness; spectral theory and functional calculi of operators; and Clifford analysis. McIntosh has been a Fellow of the Australian Academy of Science since 1986. Very recently he was awarded the Moyal Medal for 2002 for his contributions to mathematics, in particular for his fundamental contributions to harmonic analysis and partial differential equations.
ALAN McINTOSH’S PUBLICATIONS [1] Representation of bilinear forms in Hilbert space by linear operators. Trans. Amer. Math. Soc., 131:365–377, 1968. [2] On the closed graph theorem. Proc. Amer. Math. Soc., 20:397–404, 1969. [3] Bilinear forms in Hilbert space. J. Math. Mech., 19:1027–1045, 1969/1970. [4] Hermitian bilinear forms which are not semibounded. Bull. Amer. Math. Soc., 76:732–737, 1970. [5] Counterexample to a question on commutators. Proc. Amer. Math. Soc., 29:337–340, 1971. [6] On the comparability of A1/2 and A∗1/2 . Proc. Amer. Math. Soc., 32:430–434, 1972. [7] Functions and derivations of C ∗ -algebras. J. Funct. Anal., 30(2):264– 275, 1978. [8] The Toeplitz-Hausdorff theorem and ellipticity conditions. Amer. Math. Monthly, 85(6):475–477, 1978. [9] Second-order properly elliptic boundary value problems on irregular plane domains. J. Differential Equations, 34(3):361–392, 1979. [10] On representing closed accretive sesquilinear forms as (A1/2 u, A∗1/2 v). In Nonlinear partial differential equations and their applications. Coll`ege de France Seminar, Vol. III (Paris, 1980/1981), volume 70 of Res. Notes in Math., pages 252–267. Pitman, Boston, Mass., 1982. [11] with R. R.. Coifman and Y. Meyer. L’int´egrale de Cauchy d´efinit un op´erateur born´e sur L2 pour les courbes lipschitziennes. Ann. of Math. (2), 116(2):361–387, 1982. [12] with Ronald Coifman and Yves Meyer. The Hilbert transform on Lipschitz curves. In Miniconference on partial differential equations (Canberra, 1981), volume 1 of Proc. Centre Math. Anal. Austral. Nat. Univ., pages 26–69. Austral. Nat. Univ., Canberra, 1982. [13] with Rajendra Bhatia and Chandler Davis. Perturbation of spectral subspaces and solution of linear operator equations. Linear Algebra Appl., 52/53:45–67, 1983. [14] with R. R. Coifman and Y. Meyer. Estimations L2 pour les noyaux singuliers. In Conference on harmonic analysis in honor of Antoni Zygmund, Vol. I, II (Chicago, Ill., 1981), Wadsworth Math. Ser., pages 287–294. Wadsworth, Belmont, CA, 1983.
iv
v
[15] with B. Jefferies, editors. Miniconference on operator theory and partial differential equations, volume 5 of Proceedings of the Centre for Mathematical Analysis, Australian National University, Canberra, 1984. Australian National University Centre for Mathematical Analysis. [16] Square roots of operators and applications to hyperbolic PDEs. In Miniconference on operator theory and partial differential equations (Canberra, 1983), volume 5 of Proc. Centre Math. Anal. Austral. Nat. Univ., pages 124–136. Austral. Nat. Univ., Canberra, 1984. [17] Square roots of elliptic operators. J. Funct. Anal., 61(3):307–327, 1985. [18] with Yves Meyer. Alg`ebres d’op´erateurs d´efinis par des int´egrales singuli`eres. C. R. Acad. Sci. Paris S´er. I Math., 301(8):395–397, 1985. [19] with Alan Pryde. The solution of systems of operator equations using Clifford algebras. In Miniconference on linear analysis and function spaces (Canberra, 1984), volume 9 of Proc. Centre Math. Anal. Austral. Nat. Univ., pages 212–222. Austral. Nat. Univ., Canberra, 1985. [20] with Alan J. Pryde, editors. Miniconference on linear analysis and function spaces, volume 9 of Proceedings of the Centre for Mathematical Analysis, Australian National University, Canberra, 1985. Australian National University Centre for Mathematical Analysis. [21] Clifford algebras and the double-layer potential operator. In Nonlinear analysis, function spaces and applications, Vol. 3 (Litomyˇsl, 1986), volume 93 of Teubner-Texte Math., pages 69–76. Teubner, Leipzig, 1986. [22] Operators which have an H∞ functional calculus. In Miniconference on operator theory and partial differential equations (North Ryde, 1986), volume 14 of Proc. Centre Math. Anal. Austral. Nat. Univ., pages 210– 231. Austral. Nat. Univ., Canberra, 1986. [23] When are singular integral operators bounded? In Miniconference on geometry and partial differential equations (Canberra, 1985), volume 10 of Proc. Centre Math. Anal. Austral. Nat. Univ., pages 141–149. Austral. Nat. Univ., Canberra, 1986. [24] with Brian Jefferies and Werner Ricker, editors. Miniconference on operator theory and partial differential equations, volume 14 of Proceedings of the Centre for Mathematical Analysis, Australian National University, Canberra, 1986. Australian National University Centre for Mathematical Analysis. [25] with Alan Pryde. A functional calculus for several commuting operators. Indiana Univ. Math. J., 36(2):421–439, 1987. [26] with Tao Qian. Fourier theory on Lipschitz curves. In Miniconference on harmonic analysis and operator algebras (Canberra, 1987), volume 15 of Proc. Centre Math. Anal. Austral. Nat. Univ., pages 157–166. Austral. Nat. Univ., Canberra, 1987. [27] with A. Pryde and W. Ricker. Estimates for solutions of the operator P A equation m j=1 j QBj = U . In Special classes of linear operators and other topics (Bucharest, 1986), volume 28 of Oper. Theory Adv. Appl., pages 197–207. Birkh¨auser, Basel, 1988.
vi
[28] with Alan Pryde and Werner Ricker. Comparison of joint spectra for certain classes of commuting operators. Studia Math., 88(1):23–36, 1988. [29] with Alan Pryde and Werner Ricker. Systems of operator equations and perturbation of spectral subspaces of commuting operators. Michigan Math. J., 35(1):43–65, 1988. [30] Clifford algebras and the higher-dimensional Cauchy integral. In Approximation and function spaces (Warsaw, 1986), volume 22 of Banach Center Publ., pages 253–267. PWN, Warsaw, 1989. [31] The square root problem for elliptic operators: a survey. In Functionalanalytic methods for partial differential equations (Tokyo, 1989), volume 1450 of Lecture Notes in Math., pages 122–140. Springer, Berlin, 1990. [32] with Ian Doust, Brian Jefferies, and Chun Li, editors. Miniconference on Operators in Analysis, volume 24 of Proceedings of the Centre for Mathematical Analysis, Australian National University, Canberra, 1990. Australian National University Centre for Mathematical Analysis. [33] with T. Qian. Singular integrals along Lipschitz curves with holomorphic kernels. Approx. Theory Appl., 6(4):40–57, 1990. [34] with Atsushi Yagi. Operators of type ω without a bounded H∞ functional calculus. In Miniconference on Operators in Analysis (Sydney, 1989), volume 24 of Proc. Centre Math. Anal. Austral. Nat. Univ., pages 159–172. Austral. Nat. Univ., Canberra, 1990. [35] with Tao Qian. Convolution singular integral operators on Lipschitz curves. In Harmonic analysis (Tianjin, 1988), volume 1494 of Lecture Notes in Math., pages 142–162. Springer, Berlin, 1991. [36] with Chun Li and Stephen Semmes. Convolution singular integrals on Lipschitz surfaces. J. Amer. Math. Soc., 5(3):455–481, 1992. [37] with Tao Qian. Fourier multipliers on Lipschitz curves. Trans. Amer. Math. Soc., 333(1):157–176, 1992. [38] with Pascal Auscher and Philippe Tchamitchian. Noyau de la chaleur d’op´erateurs elliptiques complexes. Math. Res. Lett., 1(1):35–43, 1994. [39] with Chun Li and Tao Qian. Clifford algebras, Fourier transforms and singular convolution operators on Lipschitz surfaces. Rev. Mat. Iberoamericana, 10(3):665–721, 1994. [40] Review of Clifford Algebra and Spinor-Valued Functions, a Function Theory for the Dirac Operator, R. Delanghe, F. Sommen and V. Souˇcek. Bull. Amer. Math. Soc., 32:344–348, 1995. [41] Clifford algebras, Fourier theory, singular integrals, and harmonic functions on Lipschitz domains. In Clifford algebras in analysis and related topics (Fayetteville, AR, 1993), Stud. Adv. Math., pages 33–87. CRC, Boca Raton, FL, 1996. [42] with David Albrecht and Xuan Duong. Operator theory and harmonic analysis. In Instructional Workshop on Analysis and Geometry, Part III (Canberra, 1995), volume 34 of Proc. Centre Math. Appl. Austral. Nat. Univ., pages 77–136. Austral. Nat. Univ., Canberra, 1996.
vii
[43] with Michael Cowling, Ian Doust, and Atsushi Yagi. Banach space operators with a bounded H ∞ functional calculus. J. Austral. Math. Soc. Ser. A, 60(1):51–89, 1996. [44] with Xuan Thinh Duong. Functional calculi of second-order elliptic partial differential operators with bounded measurable coefficients. J. Geom. Anal., 6(2):181–205, 1996. [45] with Chun Li. Clifford algebras and H∞ functional calculi of commuting operators. In Clifford algebras in analysis and related topics (Fayetteville, AR, 1993), Stud. Adv. Math., pages 89–101. CRC, Boca Raton, FL, 1996. [46] with Dorina Mitrea and Marius Mitrea. Relich type identities for onesided monogenic functions in Lipschitz domains and applications. In Proceedings of the Symposium on Analytical and Numerical Methods in Quaternionic and Clifford Analysis (Seiffen, Germany, 1996), pages 135–143. 1996. [47] with Pascal Auscher and Andrea Nahmod. Holomorphic functional calculi of operators, quadratic estimates and interpolation. Indiana Univ. Math. J., 46(2):375–403, 1997. [48] with Pascal Auscher and Andrea Nahmod. The square root problem of Kato in one dimension, and first order elliptic systems. Indiana Univ. Math. J., 46(3):659–695, 1997. [49] with C. Li, K. Zhang, and Z. Wu. Compensated compactness, paracommutators, and Hardy spaces. J. Funct. Anal., 150(2):289–306, 1997. [50] with David Albrecht and Edwin Franks. Holomorphic functional calculi and sums of commuting operators. Bull. Austral. Math. Soc., 58(2):291– 305, 1998. [51] with Pascal Auscher and Philippe Tchamitchian. Heat kernels of second order complex elliptic operators and applications. J. Funct. Anal., 152(1):22–73, 1998. [52] with Edwin Franks. Discrete quadratic estimates and holomorphic functional calculi in Banach spaces. Bull. Austral. Math. Soc., 58(2):271–290, 1998. [53] with Brian Jefferies. The Weyl calculus and Clifford analysis. Bull. Austral. Math. Soc., 57(2):329–341, 1998. [54] with Xuan Thinh Duong. Singular integral operators with non-smooth kernels on irregular domains. Rev. Mat. Iberoamericana, 15(2):233–265, 1999. [55] with Xuan Thinh Duong. The Lp boundedness of Riesz transforms associated with divergence form operators. In Workshop on Analysis and Applications (Brisbane, 1997), volume 37 of Proc. Centre. Math. Anal. Austral. Nat. Univ., pages 15–25. Austral. Nat. Univ., Canberra, 1999. [56] with Brian Jefferies and James Picton-Warlow. The monogenic functional calculus. Studia Math., 136(2):99–119, 1999. [57] with Marius Mitrea. Clifford algebras and Maxwell’s equations in Lipschitz domains. Math. Methods Appl. Sci., 22(18):1599–1620, 1999.
viii
[58] with Jeff Hogan, Chun Li, and Kewei Zhang. Global higher integrability of Jacobians on bounded domains. Ann. Inst. H. Poincar´e Anal. Non Lin´eaire, 17(2):193–217, 2000. [59] with Andrea Nahmod. Heat kernel estimates and functional calculi of −b∆. Math. Scand., 87(2):287–319, 2000. [60] with Pascal Auscher, Steve Hofmann, Michael Lacey, John Lewis, and Philippe Tchamitchian. The solution of Kato’s conjectures. C. R. Acad. Sci. Paris S´er. I Math., 332(7):601–606, 2001. [61] with Pascal Auscher, Steve Hofmann, and Philippe Tchamitchian. The Kato square root problem for higher order elliptic operators and systems on Rn . J. Evol. Equ., 1(4):361–385, 2001. Dedicated to the memory of Tosio Kato. [62] with Andreas Axelsson, Ren´e Grognard, and Jeff Hogan. Harmonic analysis of Dirac operators on Lipschitz domains. In Clifford analysis and its applications (Prague, 2000), volume 25 of NATO Sci. Ser. II Math. Phys. Chem., pages 231–246. Kluwer Acad. Publ., Dordrecht, 2001. [63] with Alexander Isaev, Andrew Hassell, and Adam Sikora, editors. Geometric analysis and applications, volume 39 of Proceedings of the Centre for Mathematics and its Applications, Australian National University, Canberra, 2001. Australian National University Centre for Mathematics and its Applications. [64] with Steve Hofmann. The solution of the Kato problem in two dimensions. In Proceedings of the 6th International Conference on Harmonic Analysis and Partial Differential Equations (El Escorial, Spain, 2000), pages 143–160, Publicacions Matem`atiques, Univ. Aut`onoma de Barcelona, 2002. [65] with Steve Hofmann and Michael Lacey. The solution of the Kato problem for divergence form elliptic operators with Gaussian heat kernel bounds. Ann. of Math. (2), 156(2):623–631, 2002. [66] with Pascal Auscher, Steve Hofmann, Michael Lacey, and Ph. Tchamitchian. The solution of the Kato square root problem for second order elliptic operators on Rn . Ann. of Math. (2), 156(2):633–654, 2002. [67] with Zengjian Lou. Divergence-free Hardy space on Rn+ . Sci. China Ser. A, to appear.
ON A UNIFORM APPROACH TO SINGULAR INTEGRAL OPERATORS SERGEY S. AJIEV Abstract. The essence of an approach to the boundedness of singular integral operators based on several parameterized classes of new conditions one of which includes, in particular, H¨ormander condition and its known variations is exposed. Most of the attention is paid to the comparison with known results in the same settings. The feature of dependence from some parameters being integer is revealed. As some of applications, the existence of functional calculus and variants of Littlewood-Paley-type decompositions in its terms without any requirements of smoothness or absolute value bounds of the kernels of the corresponding holomorphic semigroups is shown.
1. Introduction The main goal of this note is to display the idea of a unified point of view on sufficient conditions formulated in the style of the H¨ormander one for the boundedness of singular integral operators and to motivate its usefullness by means of comparison with known close results (including the theory of Calder´on-Zygmund operators). In particular, the here introduced AD -classes of singular integral operators extend and generalize Calder´on-Zygmund operators and closely related operators possessing H ∞ -calculus. In line of this main purpose, formulations of assertions under consideration are partly included also in more general forms. We consider also the definitions of AD -classes in reduced forms while the complete ones are represented in [22]. The same source contains results on boundedness of singular integral operators (SIO) from one smooth function space to another (corresponding to the ”upper case” in the sense of Section 3 below) not included in this note too. The theory of singular integral operators has a half-century background of intensive development. The main ingredient — decomposition lemma — appeared in 1952 thanks to A.P. Calder´on and A. Zygmund (see [6]). In 1960, L. H¨ormander (see [14])introduced his This work was supported by Russian Fund for Basic Research (grant No. 02-0100602) and INTAS (grant No. 99-01080). 1
2
SERGEY S. AJIEV
“Cancellation condition”, and, since that time, the notion of singular integral operator (SIO) is understood as follows. A SIO T is an integral operator defined, in a sense, by means of the kernel K , s.t. T ∈ L(Lp0 , Lp0 ) : Z T f (x) = p.v. K(x, y)f (y)dy. The H¨ormander condition states that for any y, z ∈ Rn , and some C>0 Z |K(x, y) − K(x, z)|dx ≤ CH < +∞. (class H) |x−z|≥2|y−z|
Every SIO satisfying (H) (from the class H ) is bounded (admits an extention) from Lp to Lp , 1 < p < ∞ , from H1 to L1 and from L1 to L1,∞ (weak- L1 ). In addition, the adjoint operator is bounded from L∞ to BM O . Presented in this note results with the same statements as just mentioned are discussed in the section 3. One can point out that another condition weaker then H was presented by X. Duong and A. Mc Intosh (1999) in [9], and our approach permits to weaken it in the same settings (see AAD -condition in [22]). Nowadays the following definition of Calder´on-Zygmund operator (CZO) is the most commonly accepted. A CZO is a SIO T satisfying for some 0 < δ < 1 : a) |K(x, y)| < C/|x − y|n ; b) |K(x, y) − K(x, z)| ≤ C|y − z|δ |x − z|−(n+δ) for |x − z| ≥ 2|y − z| We are not imposing absolute value conditions like a) at all but one of the introduced here AD -classess contains conditions which are equivalent , or weaker then the above mentioned ones. Namely, ADx (L1 , ∞, 0, 0, 0) is equivalent to the H¨ormander integral condition, and ADx (L∞ , ∞, δ, δ, δ) in def. 2.5 is weaker then property b) of Calder´on-Zygmund operator with another one. In 1972, C.L. Fefferman, E.M. Stein (see [12]) proved (particularly) H1 − L1 and L∞ − BM O -boundedness of Calder´on-Zygmund operators. Let us pay more attention to the Hp -theory of SIOs. R. Coifman (see [7]) obtained (1974) Hp − Hp boundedness of CZO for the case of one dimension. In 1986, J. Alvarez and M. Milman (see [3]) established Hp − Hp -boundedness excluding the limiting cases (i.e. p > n/(n + δ), 0 < δ < 1 ). More precisely their assertion reads as follows: a CZO T satisfying the orthogonality condition T ? P0 = 0 is
SINGULAR INTEGRAL OPERATORS
3
bounded on Hp . Here the orthogonality condition T ? PN = 0 means R α x T a = 0 for any |α| ≤ N, and a ∈ C0∞ orthogonal to PN . Extension of this result to 0 < p < 1 was pointed out by several authors: δ -CZO satisfying T ? P[δ] is bounded on Hp for 0 < p ≤ 1 . Next we recall the definition of δ -CZO. Let s = 1 for δ ∈ N , and s = {δ} otherwise. Let T be a SIO, then it is δ -CZO if it satisfies a) |K(x, y)| < C/|x − y|n ; b) |Dyα K(x, y) − Dyα K(x, z)| ≤ C|y − z|s |x − z|−(n+|α|+s) for |x − z| ≥ 2|y − z|, |α| = [δ] . Similarly to the case of CZOs, some of the presented AD -conditions (for example, ADx (∞, L∞ , l∞ , δ, δ, δ) ) in this note are weaker then condition b) of the δ -CZOs. We provide a direct analog of the H¨ormander condition in this case too (e.g. ADx (u, Lq , l∞ , δ, δ, δ), u, q ∈ [1, ∞) ). One should add that J. Alvarez (1992) (see [2]) showed the lack of Hp − Lp (and Hp − Hp ) boundedness for p = n/(n + δ) . In 1994, D. Fan (see [10]), exploiting Littlewood-Paley-theory approach, considered the limiting case for a convolution δ -CZO T , that is, he demonstrated that, under the above conditions, T is bounded from Hp to Hp,∞ . R. Fefferman and F. Soria (1987) (see [13]) proved H1,∞ − L1,∞ boundedness for a convolution SIO T satisfying the following Dini condition: Z Z 1/2 |K(x − h) − K(x)|dx. Γ(t)dt/t < ∞, where Γ(t) = sup h6=0
0
|x|>2|h|/δ
In 1988 (publ. 1991), using a similar approach, H. Liu (see [15]) investigated boundedness properties of a convolution SIO (in particular, a CZO) in the setting of homogeneous groups and obtained the following results: a) Hp = Hp,∞ -boundedness for CZO (without condition a) ), if n/(n + 1) < p < 1 ; b) Hp,∞ − Lp,∞ -boundedness, if n/(n + 1) < p ≤ 1 , for SIO T satisfying Z 1/2 Γ(t)p | log t|tnp−1−n dt < ∞, where 0 Z |K(x − h) − K(x)|dx; Γ(t) = sup h6=0
|h|/δ<|x|<4|h|/δ
c) Hp,∞ − Hp,∞ -boundedness, if n/(n + 1) < p < 1 , for a ω -CZO (without a) -cond.), that is for a SIO T satisfying |K(x − y) − K(x)| < C|x|−n ω(|y|/|x|), |x| > 2|y|,
4
SERGEY S. AJIEV
where ω is a nondecreasing function with Z
1/2 n−n/p−1
t
| log t|2/p+ε ω(t)dt < +∞ for some ε > 0.
0
But, for ω(t) = ts , one needs s > n/p − n , i.e. a nonlimiting case. The theorems in the 5th section contain extensions or additions (including the case 0 < p ≤ n/(n+1) ) to most of these results concerning the Hp -theory (the additions to the rest and complete proofs are in [22]). It is interesting to point out the “off-diagonal” case of the Calder´onZygmund-H¨ormander result on boundedness of a SIO proved in 1961 by J.T. Schwartz (see [20] and an extension due to H. Triebel [21]): for 1 < p0 ≤ r0 < ∞, 1 ≤ q , 1 + 1/r0 = 1/p0 + 1/q , suppose that a convolution operator T with kernel K is bounded from Lp to Lr Z and satisfies |K(x − y) − K(x)|q dx ≤ C. |x|>2|y|
Then T is bounded from Lp to Lr for 1 < p ≤ p0 , 1+1/r = 1/p+1/q and from L1 to Lq,∞ . But this (“off-diagonal”) setting of the SIO theory has not attracted much attention since that time even in spite of the work ([17], 1963) of P.I. Lizorkin on (Lp , Lq ) -multipliers theorem. All the general forms of the assertion of this note include the “offdiagonal” case. Some other positive features of our approach to the boundedness of SIO are the following: a) cases of operators sending Hardy-Lorentz space to both Hardy-Lorentz and Lorentz space or, acting between spaces of smooth functions are covered; b) a new effect of the dependence from some parameters being integer (in particular, the case n/p ∈ N (in isotropic situation) for H -space theory) has been observed; c) in some cases parameters of the “target” space are shown to be optimal; d) the case 0 < p ≤ n/(n+1) for the H −H -boundedness does not require smoothness assumptions; e) some obtained sufficient conditions of boundedness of a SIO under consideration have stronger known analogs but the other do not; f ) the approach admits consideration an anisotropically SIO in the sense of [5](1966); g) the proofs of the results analogous to classical ones are no more complicated then their counterparts. The section 4 is devoted to some applications of the considered here and in [22] general forms of SIO boundedness assertions to questions connected with functional calculus. There we consider the class of operators with the kernels of their holomorphic semigroups satisfying Poisson-like AD -estimates introduced in this note and providing, in
SINGULAR INTEGRAL OPERATORS
5
this terms, sufficient conditions for the existence of a functional calculus of some operators in Hardy and other function spaces, extending results of D. Albrecht, X. Duong and A. Mc Intosh (1995) (see [1, 8]). For the limiting values of parameters, the norm estimate (for the bounded extension) of the form kφ(T )kL(X,Y ) ≤ CkφkH∞ is proved for X being a Hardy space and Y — a Hardy-Marcinkiewicz, so that X 6= Y . Another application is a continuous, in the sense of [19] (1972), form of Littlewood-Paley theorem in term of the above mentioned functional calculus. It should be noted that classical approach to this theorem relied on the properties of Hilbert transforms even in weighted multiple (product) case (see [18]) (1967). Instead, in this note we are following the approach of direct application of vector-valued SIO boundedness results used by O.V. Besov (1984) in [4] to extend Littlewood-Paley inequality to the Lp -spaces with mixed norm of functions periodic in some directions and by X. Duong (1990) to extend the existence of H ∞ -functional calculus on L2 to one in Lp , p 6= 2 . The author is pleased to thank O.V. Besov for formulating the general problem to extend classical H¨ormander SIO boundedness result to Hp (Lp ) spaces with p < 1 several years ago, A. Mc Intosh and A. Sikora for formulating the problem to obtain a Littlewood-Paley type characterization of Hardy spaces in terms of functional calculus and A. Mc Intosh with T. ter Elst for their comments improving the manuscript of this paper. This work was conducted in The Center for Mathematics and its Applications of The Australian National University. 2. Definitions and Designations. Assume N0 = N ∪ {0}. For a set E , n ∈ N let E n be the Cartesian product. Let A be a Banach space and denote by k · |Ak = k · kA the norm in space A . For t ∈ (0, ∞] let lt be a (quasi)normed P 1/t space of sequences with finite (quasi)norm k{α}|lt k = ( i |αi |t ) for t 6= ∞ , or k{α}|l∞ k = supi |αi | ; Assume also designation lt,log for the (quasi)normed space ofPsequences with the finite norm k{α}|lt,log k = k{β}|lt k, where βj = i≥j |αi | . For an measurable subset G of Rn , let X(G, A) be a function space of all (strongly) measurable functions f : G → A with some quasiseminorm k · |X(G, A)k . In particular, for p, q ∈ (0, ∞] let Lp,q (G, A) be the Bochner-Lebesgue-Lorentz space of
6
SERGEY S. AJIEV
all (strongly) measurable functions f : G → A with the finite norm kf |Lp,q (G, A)k = kkf kA |Lp,q (G)k . Let Q0 := [−1, 1]n , Qt (z) := t > 0, z ∈ Rn . Let Pλ (A) Pz + tQ0 for α be the space of polynomials { |α|≤λ cα x : cα ∈ A} . Definition 2.1. For u ∈ [1, ∞], t > 0, λ ≥ 0, x ∈ Rn, f ∈ L1,loc (Rn , A), we shall refer to the following local approximation functional by means of polynomials as to the D− functional: Du (t, x, f, λ, A) = t−n/u kf − Pt,x,λ f |Lu (Qt (x), A)k, where Pt,x,λ : Lu (Qt (x)) → Pλ (A) is a surjective projector. For simplicity, we shall understood Du (t, x, f, λ) to be Du (t, x, f, λ, A) , if A = R, C . If function f depends also from the two (vector) variables x, y , f = f (x, y) , and f|x=w (y) := f (w, y) then Duy (t, z, f (w, ·), λ, A) = Du (t, z, f|x=w , λ, A). Let C0∞ (G) be the space of infinitely differentiable and compactly supported in the open set G functions. Define the local maximal functional on a function f by M (t, x, f ) := sup{|t−n ϕ(·/t) ∗ f |(y) : |y − x| ≤ $t, ϕ ∈ C0∞ (Q0 )}. Definition 2.2. For p, q ∈ (0, ∞] , let Hp,q (Rn , A) be a completion of quasinormed space of locally summable A -valued functions f with a finite quasinorm
n n
kf |Hp,q (R , A)k := sup M (t, ·, f ) |Lp,q (R ) . t>0 Remark 2.3. It will be used that Hp,q (Rn , A) = Lp,q (Rn , A) for p > 1 . Throughout the article we shall deal with particular cases of the following one. For ϕ ∈ C0∞ , b > 1 , χQ0 ≤ ϕ ≤ χbQ0 , let Z Z (T f )(x) := K(x, y)f (y)dy := lim K(x, y)f (y)ϕ(ε(y − x))dy, ε→0
where be a singular integral operator. Moreover, the kernel K : Rn × Rn → R(C) is measurable and such that for almost every x ∈ Rn n the function K(x, ·) ∈ Lloc 1 (R \ {x}) . We shall also assume that the operator T is bounded from Lθ0 into Lθ1 for some θ0 , θ1 ∈ (0, ∞) , and to be in the union of the following classes. Remark 2.4. More general case of operator-valued kernels corresponding operators T defined on vector-valued functions in the settings, particularly, of the section 5 is considered in [22] and used in the section (4) .
SINGULAR INTEGRAL OPERATORS
7
Definition 2.5. Assume λ0 , λ1 , γ ∈ [−n, ∞) , u, q, q1 ∈ (0, ∞] , γ ≥ 0 , δ > 0, b > 1 . Let X := X(N0 ) be a (quasi)(semi)normed space of w sequences, Eq,q (Rn ) be the weighted Lorentz space with the norm 1 ,λ1
w n λ1 +n/q 0 n kf |Eq,q (R )k = f (·)| · −w| | L (R )
, q,q 1 1 λ1 and ∆i (r, w) = Qδrbi+1 (w)\Qδrbi (w), i ∈ N . Then it will be understood that: a) T ∈ ADx (u, Lq , X, λ0 , λ1 , γ) , or b) T ∈ ADx (Lq , u, X, λ0 , λ1 , γ) , if, correspondingly, the sequence w µi (r, w) := kr−λ0 Duy (r, w, K(·, y)χ∆i (·), γ)|Eq,q,λ (Rn )k, i ∈ N0 , or 1
w µi (r, w) := r−λ0 Duy (r, w, K(·, y)χ∆i (·), γ, Eq,q,λ (Rn )), i ∈ N0 , 1
is bounded in X by a constant C > 0 uniformly in r > 0, w ∈ Rn ; c) T ∈ ADx (u, Lq,q1 , λ0 , λ1 , γ) , or d) T ∈ ADx (Lq,q1 , u, λ0 , λ1 , γ) , if, correspondingly, the function 0
µ(r, w) := kr−λ0 Duy (r, w, K(·, y)| · −w|λ1 +n/q , γ)|Lq,q1 (Rn \ Qrδ (w))k, or w µ(r, w) := r−λ0 Duy (r, w, K(·, y)χRn \Qrδ (w) (·), γ, Eq,q (Rn )), 1 λ1 is bounded by a constant C > 0 uniformly in r > 0, w ∈ Rn . The infimum of constants C in each case will be designated by means of CAD for the corresponding AD -condition.
Remark 2.6. It can be noted that the definitions of AD -classes have and equivalent continuous forms, which means also their independence from the parameter b > 1 . Definition 2.7. Let γ0 , γ1 R≥ 0 . An operator T will be assumed form ∞ the class ORT R x (γ0 , γ1 ) if πT φ = 0 for all π ∈ Pγ1 and φ ∈ C0 , such that φπ = 0 for each π ∈ Pγ0 . Definition 2.8. For Ω ⊂ C let {T (z)}z∈Ω be a family of integral operators with the corresponding C -valued kernels {Kz }z∈Ω , Kz = Kz (x, y) , x, y ∈ Rn . We shall assume that the family {T (z)}z∈Ω satisfies Poisson-type ADx -estimates with parameters u ∈ [1, ∞], λ ≥ 0 on the domain Ω if for some , m ∈ (0, ∞) and any w, x ∈ Rn , z ∈ Ω, r ∈ (0, ∞) λ −(n+λ+) r |x − w| −mn Du (r, w, Kz (x, ·), λ) ≤ C |z| 1+ , |z|m |z|m T (z) ∈ ORTx (λ, λ). And Kz (x, y) is understood satisfying Poisson-type ADy -estimate if KzI (x, y) = Kz (y, x) satisfies Poisson-type ADx -estimate.
8
SERGEY S. AJIEV
Definition 2.9. We shall understood operator T defined by the kernel K(x, y) to be in ADy -class, or ORTy (γ0 , γ1 ) -class if, and only if, the corresponding operator T I defined by the kernel K I (x, y) = K(y, x) is in the corresponding ADx -class, or, correspondingly, ORTx (γ0 , γ1 ) class. 3. Counterparts of Known (Clasical) Results. For the sake of simplicity, we shall only consider in this section AD classes with q = 1 , X = l1 and λ0 = λ1 = γ = 0 . Remark 3.1. In spite of the relation ADx (u, L1 , 0, 0, 0) = ADx (u, L1 , l1 , 0, 0, 0) ⊂ ⊂ ADx (L1 , u, l1 , 0, 0, 0) ⊂ ADx (L1 , u, 0, 0, 0), not coinciding AD -classes will be discussed in the proofs separately to demonstrate the approach in more general cases. Let us point out that the class of operators satisfying H¨ormander condition H is equal to the class ADx (L1 , ∞, l1 , 0, 0, 0) . We shall show the inclusion H ⊂ ADx (L1 , ∞, l1 , 0, 0, 0) . The opposite one was pointed out to the author by A. Mc Intosh . Indeed, the corresponding kernels should have a uniformly bounded for any z ∈ Rn , r > 0 quantity Z A(r, z) = inf
|K(x, y) − c(x)|dx.
sup
c(x) {y:|y−z|≤r}
|x−z|≥2r
And, supposing, for fixed z, r , c(x) to be equal to K(x, z) , we can note that Z A(r, z) ≤ sup |K(x, y) − K(x, z)|dx ≤ {y:|y−z|≤r}
|x−z|≥2r
Z ≤ sup y
|K(x, y) − K(x, z)|dx ≤ CH < +∞, |x−z|≥2|y−z|
where CH is the constant in H¨ormander condition (class H ). 3.1. Lower “Summability” Case. Theorem 3.2. For p0 ∈ (1, ∞] , u ∈ [1, ∞] , let T be a SIO from ADx (L1 , u, l1 , 0, 0, 0) ∪ ADx (L1 , u, 0, 0, 0) ∪ ADx (u, L1 , 0, 0, 0) , bounded from Lp0 into itself. Then, a) T ∈ L(Lp,q (Rn )) for p ∈ (1, p0 ] , q ∈ (0, ∞] ; b) T ∈ L(L1 (Rn ), L1,∞ (Rn )) if u = ∞ ; c) T ∈ L(H1 (Rn ), L1 (Rn )) .
SINGULAR INTEGRAL OPERATORS
9
Proof of the Theorem 3.2. We suppose that kernel K(x, y) corresponds to the operator T . One should note that part a) of the theorem is a consequence of both b) and c) in view of the interpolation properties of the scale of Hardy-Lebesgue spaces (see [11]). To the first, let us recall that the statements of the parts b) and c) are implied as in classical approach) by the estimate Z |T a|dx ≤ CAD , (1) Rn \Qrδ (z)
where r > 0, z ∈ Rn and a is a (1, L∞ , 0) -, or a (1, 1, 0) -atom in the case of the part c) , or b) correspondingly. Indeed, in the case c) , the atomic decomposition result for H1 (see [7, 16]) permits us to prove the boundedness of T on (1, ∞, 0) -atoms only, what follows from (1) and Z 0
0
|T a|dx ≤ (rδ)n/p0 kT akp0 ≤ δ n/p0 kT |L(Lp0 )k,
Qrδ (z)
where a is an (1, ∞, 0) -atom. In the case b) , for a function f ∈ L1 and λ > 0 , S Calder´on-Zygmund decomposition of a set Ωλ := {x : the set {δQi } possess finite intersection M f > λ} = i∈N Qi , where P property and C|Ωλ | ≥ i |Qi | , provides representation X f = f0 + Cλ |Qi |ai , where ai is a (1, 1, 0) − atom (2) i
and kf0 |L∞ k ≤ Cλ . Therefore, Chebyshev inequality, (1) and just mentioned properties imply λ{|T f0 | > cλ} ≤ Cλ1−p0 kT f0 kpp00 ≤ CkT |L(Lp0 )kp0 kf |L1 k, λ{|T f1 | > cλ} ≤ Cλ | ∪i δQi | +
X
Z |Qi |
|T ai | ≤
Rn \δQi
≤ Cλ|Ωλ | ≤ Ckf |L1 k.
(3)
To obtain the formula (1) suppose g(x) to be a function minimizing functionals Z inf |K(x, y) − c|u dy (4) c
Qr (z)
at a.e. x if T ∈ ADx (u, L1 , l1 , 0, 0, 0) , or minimizing the functional Z u Z |K(x, y) − c(x)|dx dy, (5) Qr (z)
Rn \Qrδ (z)
10
SERGEY S. AJIEV
P if T ∈ ADx (L1,1 , u, 0, 0, 0) , or g(x) = i gi (x) , where functions {gi (x)}i∈N to minimize functionals !u !1/u Z Z inf |K(x, y) − c(x)|dx dy (6) c(x)
Qr (z)
Qr2i δ (z)\Qr2i−1 δ (z)
correspondingly if T ∈ ADy (L1 , u, l1 , 0, 0, 0) . In view of Minkowski inequality, it follows from (4) , (5) , or (6) that, correspondingly, for an arbitrarily (1, 1, 0) -(part b) ), or (1, ∞, 0) -atom with support Qr (z) , one has due to the H¨older and Minkowski inequalities, Fubini theorem and the orthogonality of atom a to constants: Z |T a|dx ≤ Rn \Qrδ (z)
Z (K(x, y) − g(x))a(y)dy dx ≤ ≤Q= Rn \Qrδ (z) Z 1/u Z −n/u u ≤ r inf |K(x, y) − c| dy dx ≤ CAD , c Rn \Qrδ (z) Qr (z) Z Z Q≤ |K(x, y) − g(x)|dx|a(y)|dy ≤ Z
Rn \Qrδ (z)
Qr (z)
≤ inf r
−n/u
Z
c
Q≤
|K(x, y) − c(x)|dx
≤
X i∈N
inf r−n/u c
|K(x, y) − gi (x)|dx|a(y)|dy Q2i rδ (z)\Q2(i−1) rδ (z)
≤
!1/u
!u
Z |K(x, y) − c(x)|dx
Qr (z)
(8)
Z
Qr (z)
Z
≤ CAD ,
dy
Rn \Qrδ (z)
XZ i∈N
1/u
u
Z
Qr (z)
(7)
dy
≤ CAD . (9)
Q2i rδ (z)\Q2(i−1) rδ (z)
In this manner, estimates (7 − 9) motivate (1) . 3.2. Upper “Summability” Case. The next theorem can be derived from the previous one by means of duality considerations but such approach will not work definitely, in particular, in the case of vectorvalued functions, or will require additional duality results to consider scales other than H1 − Lp − BM O . Thus, proof provided does not rely on duality. Theorem 3.3. For p0 ∈ (1, ∞] , u ∈ [1, ∞) , let T be a SIO from ADy (L1 , u, l1 , 0, 0, 0) ∪ ADy (L1 , u, 0, 0, 0) ∪ ADy (u, L1 , 0, 0, 0) , bounded from Lp0 into itself. Then, a) T ∈ L(Lp,q (Rn )) for p ∈ [p0 , ∞) , q ∈ (0, ∞] ; b) T ∈ L(L∞ (Rn ), BM O(Rn )) .
SINGULAR INTEGRAL OPERATORS
11
Proof of the Theorem 3.3. We suppose that kernel K(x, y) corresponds to the operator T . One should note that we need to prove the part b) only because the part a) follows from it with the aid of the real interpolation method. Let us fix Qr (z) ⊂ Rn , f ∈ L∞ and use representation f = f0 + f1 , f0 = χQrδ (z) , where δ is a constant in the definition of the corresponding ADy -classes. Then the definition of D -functional and Lp0 boundedness of T and restriction operator f −→ f0 imply Dp0 (r, z, T f0 , 0) ≤ r−n/p0 kT f0 |Lp0 (Rn )k ≤ r−n/p0 kT |L(Lp0 )k× ×kf0 |Lp0 (Rn )k ≤ kT |L(Lp0 )kkf0 |L∞ (Rn )k ≤ kT |L(Lp0 )kkf |L∞ (Rn )k. (1) Now suppose g(y) to be a function minimizing functionals Z inf |K(x, y) − c|u dx (2) c
Qr (z)
at a.e. y if T ∈ ADy (u, L1 , l1 , 0, 0, 0) , or minimizing the functional Z u Z |K(x, y) − c(y)|dy dx, (3) Rn \Qrδ (z)
Qr (z)
P if T ∈ ADy (L1,1 , u, 0, 0, 0) , or g(y) = i gi (y) , where the functions {gi (y)}i∈N to minimize functionals !u !1/u Z Z inf |K(x, y) − c(y)|dy dx (4) c(y)
Qr (z)
Qr2i δ (z)\Qr2i−1 δ (z)
correspondingly if T ∈ ADy (L1 , u, l1 , 0, 0, 0) . In view of Minkowski inequality, it follows from (2) , or (3) , or (4) that, correspondingly, u 1/u Z Z −n/u = Du (r, z, T f1 , 0) ≤ r |(K(x, y) − g(y))f1 (y)|dy) dx Rn \Qrδ (z)
Z
Z
1/u
u
=Q≤
|K(x, y) − g(y)| dx Rn \Qrδ (z)
kf |L∞ dy, or
(5)
Qr (z)
1/u
u
Z Z Q≤
|K(x, y) − g(y)|dy
dx
, or
(6)
Rn \Qrδ (z)
X Z Q≤ i
Qr (z)
!1/u
!u
Z |K(x, y) − c(y)|dy
dx
kf |L∞ k.
Qr2i δ (z)\Qr2i−1 δ (z)
Eventually formulas (1) and one of (5, 6, 7) imply D1 (r, z, T f, 0) ≤ (CAD + kT |L(Lp0 )k)kf |L∞ k.
(7)
12
SERGEY S. AJIEV
4. Applications 4.1. Functional Calculus and Littlewood-Paley-type Theorems. In this section we have A = B = C . All the definitions and notations regarding functional calculus are understood as in the article [1] due to David Albrecht, Xuan Duong and Alan Mc Intosh , including the following definitions. 0 For 0 ≤ ω < µ < π let Sω+ := {z ∈ C| | arg z| ≤ ω} ∪ {0} , Sµ+ := 0 {z ∈ C| | arg z| < µ} , space H(Sµ+ ) be the space of all holomorphic 0 0 functions on Sµ+ endowed with the L∞ (Sµ+ ) -norm, containing sub0 0 space Ψ(Sµ+ ) := {ψ|ψ ∈ H(Sµ+ ), ∃s > 0, |ψ(z)| ≤ C|z|s (1 + |z|2s )−1 } . A closed in L2 (Rn ) operator T is said to be of type Sω+ if σ(T ) ⊂ Sω+ and for any µ > ω there exist Cµ such that |z|k(T − zI)−1 |L(L2 (Rn ), L2 (Rn ))k ≤ Cµ , z 6∈ Sω+ . Theorem 4.1. Assume q ∈ (0, ∞] , r ∈ (0, ∞]n . Let T be a oneone operator of type Sω+ in L2 (Rn ) , ω ∈ [0, π/2) , having a bounded 0 functional calculus in L2 (Rn ) for all f ∈ H∞ (Sµ+ ) for some µ > ω . Assume that for some λ, m, > 0 , [v0 , µ] ∈ (ω, π/2) and all z ∈ 0 S(π/2−µ) , the kernel Kz (x, y) of holomorphic semigroup e−zT associated with T satisfies: a) Poisson-type ADx -estimate with u = 1 ; b) Poisson-type ADy -estimate with u = 1 ; c) both Poisson-type ADx - and ADy -estimates with u = 2 . Then, correspondingly, for f ∈ H∞ (Sν0 ) for all ν > µ , f (T ) can be extended to be in: [ a) L(Hp,q (Rn ), Hp,q (Rn ))with kf (T )|L(Hp,q , Hp,q )k ≤ Ckf |L∞ k and p∈((1+λ/n)−1 ,2]
L(Hp0 (Rn ), Hp0 ,∞ (Rn ))with kf (T )|L(Hp0 (Rn ), Hp0 ,∞ (Rn ))k ≤ Ckf |L∞ k for p0 = (1 + λ/n)−1 , λ 6∈ Z ; b)
[
L(Lp,q (Rn ), Lp,q (Rn )) with kf (T )|L(Lp,q , Lp,q )k ≤ Ckf |L∞ k
p∈[2,∞)
and
[
L(X γ , X γ ) with kf (T )|L(X γ , X γ )k ≤ Ckf |L∞ k;
γ∈[0,λ),T ∈ORT y (γ,γ)
γ (Rn ) ; for X γ ∈ bγr,q (Rn ), lr,q R c) if, in addition, functional Ψν0 (f ) = Γν (π−ν0 )/2
0
(ν0 −π)/2
|f (ζ)| |dζ| |ζ|
is finite for
some Γν0 = Θ(t)te + (Θ(t) − 1)te , t ∈ R , then the following Littlewood-Paley-type estimates is true: for γ ∈ [0, λ) , T ∈
SINGULAR INTEGRAL OPERATORS
13
ORT y (γ, γ) , p ∈ ((1 + λ/n)−1 , ∞) and γ X(Rn ) ∈ Hp (Rn ), bγ∞,∞ (Rn ), bγr,q (Rn ), lr,q (Rn )
f (tT )g| X(Rn , L2, dt (R+ )) (kf |L∞ k + Ψν0 (f )) kg | X(Rn )k . t
Partial proof of the Theorem 4.1.We shall discuss only the proofs of the assertions concerning Hardy-Lorentz spaces (”Lower case”). Full proof contained in [22]. Theorem D from [1] supplies an opportunity to consider only func0 tions f from the class Ψ(Sµ+ ) in all the parts of the theorem thanks to some limiting procedure. By the theorem 5.1 and the existence of H ∞ -calculus of operator T in L2 , it is sufficient to estimate only the constants CAD from the definitions of the appropriate AD -classes. For this purpose we shall use the following representation of the operator f (T ) and its kernel, obtained by Xuan Duong (see [8]): Z Z −zT f (T ) = e n(z)dz, n(z) = ezζ f (ζ)dζ, (1) Γν0
Γ0
Γν0 = Θ(t)te(π−ν0 )/2 +(Θ(t)−1)te(ν0 −π)/2 , Γ0 = Θ(t)teν0 +(Θ(t)−1)te−ν0 , Z t ∈ R, Kf (x, y) = kz (x, y)n(z)dz, |n(z)| ≤ c|z|−1 kf k∞ . Γ0
Now using subadditivity of D -functional we have in the case a) for a function f with kf |L∞ k ≤ 1 λ −(n+λ+) Z r |x − w| −mn Du (r, w, Kf (x, ·), λ) ≤ C |z| 1+ |dz|/|z| ≤ m |z|m Γν0 |z| ≤ Crλ |x − w|−(n+λ) . Hence f (T ) ∈ ADx (1, L∞ , l∞ , λ, λ, λ) and
(2)
ADx (u, L∞ , lt,log , γ, γ, λ), γ ∈ [0, λ), t ∈ (0, ∞]. uniformly by f . It means the desirable estimate CAD (f ) ≤ Ckf |L∞ k . It is left to apply the part c) of the theorem 5.1. To prove part c) let us fix a (nonzero) function f kf |H∞ k + Ψν0 (f ) ≤ 1 , satisfying the conditions of c) , and such that for z ∈ Γ0 Z ∞ dt f 2 (tz) = cI > 0. (3) t 0 Now we can define operators Λ : g(x) → {(f (tT )g)(x)}t∈R+ , Z dt −1 −1 Λ : {h(t, x)}t∈R+ → CI f (tT )h(t, x) , t t∈R+
14
SERGEY S. AJIEV
which define an isomorphism between L2 (Rn ) and L2 (Rn , L2, dt (R+ )) t because of the theorem F from [1]. Hence, analogously to the derivation of the formula (2) , subadditivity of the D -functional, Minkowski inequality and finiteness of Ψν0 (f ) imply both for kf = Kf and for kf = KfI Z λ
r|z · |−m |z · |−mn × Du (r, w, kf (x, ·), λ, L2, dt (R+ )) ≤ C
t Γν0
× 1 + |x − w||z · |−m
−(n+λ+)
L2, dtt |n(z)| · |dz| ≤
≤ Crλ |x − w|−(n+λ) .
(4)
It means that Λ, Λ−1 ∈ ADx (1, L∞ , l∞ , λ, λ, λ) ∩ ADy (1, L∞ , l∞ , λ, λ, λ) . Thus the proof of the part c) is finished exactly as ones of the parts a), b) . 5. Examples of the Formulations of Main Results in a More General Form. In this section we shall present two theorems which are, in turn, simplified extracts from the first two main theorems in [22]. Theorem 5.1. Let T be a singular integral operator with kernel K(x, y) satisfying condition T ∈ ORTx (γ, λ1 ) , λ1 ∈ [−n, ∞) , λ0 , γ ≥ 0 and bounded from Lθ0 (Rn ) into Lθ1 (Rn ) , t, θ0 ∈ (0, ∞], θ1 ∈ [1, ∞] ; let also λ0 − λ1 = n(1/θ0 − 1/θ1 ) , pi = (1 + λi /n)−1 , i = 0, 1 , u, q ∈ [1, ∞] , q0 , q1 ∈ (0, ∞] , q > p1 , θ0 > p0 , and K(x, y) satisfies condition T ∈ ADx (u, Lq , X, λ0 , λ1 , γ) ∪ ADx (Lq , u, X, λ0 , λ1 , γ) . Then for 1/v1 − 1/v0 = 1/θ1 − 1/θ0 operator T is also bounded with its norm bounded from above by C(CAD + kT |L(Lθ0 , Lθ1 )k) in the following situations. Assuming min{1, t, p1 } ≥ q0 ≥ p0 and the space X = lt,log for λ1 ∈ Z , or X = lt for λ1 6∈ Z a) T ∈ L(Hp0 ,q0 (Rn ), Hp1 ,t (Rn )) ∩ L(Hv0 ,s (Rn ), Hv1 ,s (Rn )) for v0 ∈ (p0 , θ0 ), s ∈ (0, ∞] ; b) T ∈ L(Hv0 ,w0 (Rn ), Hv1 ,s (Rn )) for v1 < q, s ∈ [v0 , ∞] , v1 ≤ w0 ≤ min(v1 , s, 1) , and either for θ0 > 1, v0 ∈ (p0 , 1] , or θ0 ≤ 1, v0 ∈ (p0 , θ0 ) ; c) in particular, for λ0 = λT1 , θ0 = θ1 > 1 , s ∈ (0, ∞] T T ∈ L(Hp0 (Rn ), Hp0 ,t (Rn )) p∈(p0 ,1] L(Hp,s (Rn ), Hp,s (Rn )) T n n p∈(1,θ0 ] L(Lp,s (R ), Lp,s R ). Theorem 5.2. Let T be a singular integral operator with kernel K(x, y) bounded from Lθ0 (Rn ) into Lθ1 (Rn ) , t, θ0 , θ1 ∈ (0, ∞] ; let also λ0 ≥ 0 , λ1 ∈ [−n, ∞) , λ0 − λ1 = n(1/θ0 − 1/θ1 ) , finite γ ≥ 0 , pi =
SINGULAR INTEGRAL OPERATORS
15
(1 + λi /n)−1 , i = 0, 1 , u, q ∈ [1, ∞] , q ≥ p1 , θ0 > p0 . Then for 1/v1 − 1/v0 = 1/θ1 − 1/θ0 operator T ∈ ADx (u, Lq,t , λ0 , λ1 , γ) ∪ ADx (Lq,t , u, λ0 , λ1 , γ) : is also bounded with its norm bounded from above by C(CAD + kT |L(Lθ0 , Lθ1 )k) in the following situations. Assuming min{1, t, p1 } ≥ q0 ≥ p0 and K(x, y) satisfying condition a) T ∈ L(Hp0 ,q0 (Rn ), Lp1 ,t (Rn )) ∩ L(Hv0 ,s (Rn ), Lv1 ,s (Rn )) for v0 ∈ (p0 , θ0 ), s ∈ (0, ∞] ; b) T ∈ L(Hv0 ,w0 (Rn ), Lv1 ,s (Rn )) for v1 ≤ q, s ∈ [v0 , ∞] , v0 ≤ w0 ≤ min(v1 , s, 1) , and either for θ0 > 1, v0 ∈ (p0 , 1] , or θ0 ≤ 1, v0 ∈ (p0 , θ0 ) ; c) in particular, for λ0 = λ T1 , θ0 = θ1 , s ∈ (0, ∞] T ∈ L(Hp0 (Rn ), Lp0 ,t (Rn )) p∈(p0 ,θ0 ] L(Hp,s (Rn ), Lp,s (Rn )). d) Assuming u = ∞ and either v1 = p1 > 1 , or v1 = p1 = t = 1 , or under the conditions of part b) with , γ = 0 , v0 = 1 , T ∈ L(L1 (Rn , A), Lv1 ,∞ (Rn , B)), where 1 − 1/v1 = 1/θ0 − 1/θ1 . Remark 5.3. It is proved in [22] that the limiting cases of the theorems 5.1, 5.2 cannot be improved in the sense of reducing the value of the parameter t of the “target” space Hp1 ,t provide n/p1 6∈ N , or the space L(p1 , t) , correspondingly. References [1] Albrecht, D., Duong, X. and Mc Intosh , A., Operator Theory and Harmonic Analysis (Workshop in Analysis and Geometry, art III, C.M.A., A.N.U., Canberra, Jan.-Feb. 1995), 77-136, Proc. Centre Math. Appl. Austral. Nat. Univ. V. 34, Austral. Nat. Univ., Canberra, 1996. ´ [2] Alvarez, J., Hp and weak Lp continuity of Calder´on-Zygmund type operators. Fourier analysis (Orono, ME, 1992), 17-34, Lecture Notes in Pure and Appl. Math., 157, Dekker, New-York, 1994. ´ [3] Alvarez, A. J., Milman, M., H p continuity properties of Caldern-Zygmund-type operators, J. Math. Anal. Appl., V. 118 (1986), no. 1, pp. 63–79. [4] Besov, O. V., The Littlewood-Paley theorem for a mixed norm, (Russian), Studies in the theory of differentiable functions of several variables and its applications, Trudy Mat. Inst. Steklov., V. 170 (1984), pp. 31–36, 274. [5] Besov, O. V., Il’in, V. P., Lizorkin, P. I., The Lp -estimates of a certain class of non-isotropically singular integrals, engl. transl.: Soviet Math. Dokl., V. 7 (1966), pp. 1065–1069. [6] Calderon, A. P., Zygmund, A. On the existence of certain singular integrals, Acta Math., V. 88 (1952), pp. 85–139. [7] Coifman, R.R., A real variable characterization of Hp , Studia Math., V.51 (1974), pp. 269-274. [8] Duong, X. T., H∞ Functional Calculus of Elliptic Partial Differential Operators in Lp spaces, (1990), PhD Thesis, Macquary University, pp. .
16
SERGEY S. AJIEV
[9] Duong, X. and Mc Intosh , A., Singular Integral Operators with Non-smooth Kernels on Irregular Domains, Rev. Mat. Iberoamericana, V.15 (1999), No 2, pp. 233-265. [10] Fan, D. S., A weak estimate of Caldern-Zygmund operators, Integral Equations Operator Theory, V. 20 (1994), no. 4, pp. 383–394. [11] Fefferman,C., Rivi`ere, N.M., Sagher, Y., Interpolation between Hp spaces: The real method, Trans. Amer. Math. Soc. 191, (1974), pp. 75-81. [12] Fefferman,C., Stein, E.M., Hp spaces of several variables, Acta Math. V. 129, (1972), pp. 137-194. [13] Fefferman,R., Soria, F., The space Weak H1 , Studia Math., V. 85, (1987), pp. 1-16. [14] H¨ ormander, L., Estimates for translation invariant operators on Lp spaces, Acta Math., V. 104 (1960), pp. 93-139. [15] Liu, He Ping, The weak H p spaces on homogeneous groups. Harmonic analysis (Tianjin, 1988), 113–118, Lecture Notes in Math., 1494, Springer, Berlin, 1991. [16] Latter, R., A characterization of Hp (Rn ) in terms of atoms, Studia Math. V. 62, (1978), pp. 93-101. [17] Lizorkin, P. I., (Lp , Lq ) -multipliers of Fourier integrals, (Russian), Dokl. Akad. Nauk SSSR, V. 152 (1963), pp. 808–811. [18] Lizorkin, P. I., A Littlewood-Paley type theorem for multiple Fourier integrals, (Russian), Trudy Mat. Inst. Steklov, V. 89 (1967), pp. 214–230. [19] Lizorkin, P. I., A Marcinkiewicz type theorem for H -valued functions. A continuous form of the Littlewood-Paley theorem, engl. transl.: Math. USSR-Sb., V. 16 (1972), no. 2, pp. 237–243. [20] Schwartz, J. T., A remark on inequalities of Calderon-Zygmund type for vectorvalued functions, Comm. Pure Appl. Math., V. 14 (1961), pp. 785–799. [21] Triebel, H., Spaces of distributions of Besov type on Euclidean n -space. Duality, interpolation, Arkiv Mat., V. 11 (1973), pp. 13-64. [22] Ajiev, S., S.,On the boundedness of singular integral operators and other properties of some spaces of vector-valued functions, in preparation. CMA, SMS, ANU, A.C.T. 0200, Australia E-mail address:
[email protected]
ALMOST PERIODIC BEHAVIOUR OF UNBOUNDED SOLUTIONS OF DIFFERENTIAL EQUATIONS BOLIS BASIT AND A. J. PRYDE Abstract. A key result in describing the asymptotic behaviour of bounded solutions of differential equations is the classicalR result of t Bohl-Bohr: If φ : R → C is almost periodic and P φ(t) = 0 φ(s) ds is bounded then P φ is almost periodic too. In this paper we reveal a new property of almost periodic functions: If ψ(t) = tN φ(t) where φ is almost periodic and P ψ(t)/(1 + |t|)N is bounded then P φ is bounded and hence almost periodic. As a consequence of this result and a theorem of Kadets, we obtain results on the almost periodicity of the primitive of Banach space valued almost periodic functions. This allows us to resolve the asymptotic behaviour of unbounded solutions of differential equations of the form Pm (j) (t) = tN φ(t). The results are new even for scalar valb u j=0 j ued functions. The techniques include the use of reduced Beurling spectra and ergodicity for functions of polynomial growth. Keywords: almost periodic, almost automorphic, ergodic, reduced Beurling spectrum, primitive of weighted almost periodic functions, Esclangon-Landau.
1. Introduction, Notation and Preliminaries A problem arising naturally from a theorem of Bohl-Bohr-Kadets [21], (see also [4], [9, Sections 5, 6] and references therein) is to investigate the almost periodicity of the primitive P ψ when ψ = tN φ, where φ : R → X is almost periodic, X is a Banach space, and N is a non-negative integer. More generally we describe the asymptotic behaviour of solutions u : R → X of differential equations of the form Pm (j) N j=0 bj u (t) = t φ(t) where bj ∈ X and m ∈ N. We begin by introducing some notation. The function w(t) = wN (t) = (1 + |t|)N is a weight on R, satisfying in particular w(s + t) ≤ w(s)w(t). By J we will mean R, R+ or R− . A function φ : J → X is called w-bounded if φ/w is bounded and BCw (J, X) is the space of all continuous w-bounded functions, a Banach space with norm ||φ||w,∞ = supt∈R ||φ(t)|| . Following Reiter [28, p. 142], φ is w-uniformly continuous w(t) if k∆h φkw,∞ → 0 as h → 0 in J. Here ∆h φ denotes the difference 17
18
BOLIS BASIT AND A. J. PRYDE
of φ by h defined by ∆h φ(t) = φ(h + t) − φ(t). The closed subspace of BCw (J, X) consisting of all w-uniformly continuous functions is denoted BU Cw (J, X). It is not hard to show that φ is w-uniformly continuous if and only if φ/w is uniformly continuous. Furthermore, kφt+h − φt kw,∞ ≤ w(t) kφh − φkw,∞ and so (1.1)
if φ ∈ BU Cw (J, X) then the function t → φt : J → BU Cw (J, X) is continuous.
When N = 0 or equivalently w = 1 we will drop the subscript w from the names of various spaces. As an example, note that for λ, N 6= 0, the function φ(t) = tN eiλt is not bounded or uniformly continuous. However, φ is both w-bounded and w-uniformly continuous and so φ ∈ BU Cw (R, X). We define T Pw (R, X) = span{tj eiλt : 0 ≤ j ≤ N λ ∈ R} and APw (R, X) to be the closure in BU Cw (R, X) of T Pw (R, X). These are natural generalizations of the spaces T P (R, X) of Xvalued trigonometric polynomials and AP (R, X) of almost periodic functions which correspond to the case N = 0. Suppose now that u0 = ψ where u ∈ BU Cw (R, X), ψ ∈ APw (R, X) and X 6⊇ c0 , that is X does not contain a subspace isomorphically isometric to c0 . Kadets proved that necessarily u ∈ APw (R, X) when N = 0. However, the following example shows that this is not the case for general N . Indeed, we will show below that the general case is more delicate. t Example 1.1. Take X = C, N = 1, w(t) = 1+|t|, ψ(t) = w(t) cos log w(t) 1 1 and u(t) = 2 w(t) cos log w(t) + 2 w(t) sin log w(t) − sin log w(t) − 21 . Then u ∈ BU Cw (R, X), ψ ∈ APw (R, C) and u0 = ψ. However, u∈ / APw (R, C).
The proof of this assertion requires some further theory and will be given in Remark 4.4. 2. Some function spaces. In [12] a function φ : J → X is called Maak-ergodic with mean M φ = x ∈ X (see also [25], [19], [20], [11]) if for each ε > 0P there is a finite subset F ⊆ J with kRF φ − x)k < ε where RF φ = |F1 | t∈F φt . Moreover E(J, X) is the closed subspace of BC(J, X) consisting of Maak-ergodic functions and E0 (J, X) = {φ ∈ E(J, X) : M φ = 0}. If M : E(J, X) → X is the function φ → M φ, it follows that M is linear and continuous and E(J, X) = E0 (J, X) ⊕ X.
ALMOST PERIODIC BEHAVIOUR OF UNBOUNDED SOLUTIONS
19
In [12] we also defined a notion of ergodicity that applies to unbounded functions. This ergodicity differs from both that of Maak and that of Basit-G¨ unzler [9]. Indeed, the space of w-ergodic functions is defined by Ew (J, X) = { φ ∈ BCw (J, X) : φ/w ∈ E(J, X)}. Both Ew (J, X) and Ew,0 (J, X) = {φ ∈ Ew (J, X) : M (φ/w) = 0} are closed subspaces of BCw (J, X). It is convenient to introduce an even larger class. For this we need Pw (J, X) the closed subspace of BCw (J, X) consisting of polynomials on J with coefficients in X. We shall say a function φ : J → X is w -polynomially ergodic with w-mean p ∈ Pw (J, X) if (φ − p)/w ∈ E0 (J, X). The space of all such φ is denoted P Ew (J, X) and satisfies P Ew (J, X) = Ew,0 (J, X)+Pw (J, X). For N 6= 0 a w-mean is not unique and this last sum is not direct. Of course Pw (J, C) is finite dimensional and so P Ew (J, C) is a closed subspace of BCw (J, C). Moreover, we can choose a subspace PwM (J, C) of Pw (J, C) such that P Ew (J, C) = Ew,0 (J, C) ⊕ PwM (J, C). The (continuous) projection map Mw : P Ew (J, C) → PwM (J, C) then provides a unique w-polynomial mean Mw (φ) for each φ ∈ P Ew (J, C). Now set PwM (J, X) = PwM (J, C) ⊗ X and define Mw : P Ew (J, X) → PwM (J, X) P by Mw (φ) = kj=1 Mw (pj )⊗xj where φ ∈ P Ew (J, X) has w-polynomial P mean p = kj=1 pj ⊗ xj ∈ Pw (J, C) ⊗ X. Proposition 2.1. The map Mw : P Ew (J, X) → PwM (J, X) is welldefined and continuous. Moreover, for each φ ∈ P Ew (J, X), Mw (φ) is a w-polynomial mean for φ and for each of its translates. Finally, P Ew (J, X) is a closed translation invariant subspace of BCw (J, X) and P Ew (J, X) = Ew,0 (J, X) ⊕ PwM (J, X). Pk Proof. Let φ ∈ P Ew (J, X) have means p = j=1 pj ⊗ xj and q = Pm ∗ j=1 qj ⊗ yj . Then p − q ∈ Ew,0 (J, X) and so x ◦ (p − q) ∈ Ew,0 (J, C) P for all x∗ ∈ X ∗ . Hence Mw (x∗ ◦ (p − q)) = 0 = x∗ ◦ ( kj=1 Mw (pj ) ⊗ xj − Pm Pk Pm j=1 Mw (qj ) ⊗ yj ) which gives j=1 Mw (pj ) ⊗ xj = j=1 Mw (qj ) ⊗ yj showing Mw is well-defined. Also, pj − Mw (pj ) ∈ Ew,0 (J, C) and so by Lemma 2.2(a) below p − Mw (p) ∈ Ew,0 (J, X). Hence Mw (φ) is a mean for φ. Moreover, kx∗ ◦ Mw (φ)kw,∞ = kMw (x∗ ◦ φ)kw,∞ ≤ c kx∗ ◦ φkw,∞ = c supt∈J kx∗ ◦ φ(t)k /w(t) ≤ c kx∗ k kφkw,∞ . Hence, kx∗ ◦ψk
|||Mw (φ)||| ≤ c kφkw,∞ where |||ψ||| = supx∗ ∈X ∗ kx∗ kw,∞ for ψ ∈ Pw (J, X). By Lemma 2.2(b) below, Mw is continuous. If (φn ) is a sequence in P Ew (J, X) converging to φ in BCw (J, X), let pn = Mw (φn ). Then (pn ) converges to some p ∈ PwM (J, X) and so ( φnw−pn ) converges to φ−p in BC(J, X). By the continuity of the Maak mean w φ−p function, M ( w ) = 0 and so P Ew (J, X) is closed. That P Ew (J, X) =
20
BOLIS BASIT AND A. J. PRYDE
Ew,0 (J, X) + PwM (J, X) is clear and that the sum is direct follows from the Hahn-Banach theorem. Finally, for each t ∈ J we have φt −p = ∆wt φ + φ−p and so, by Lemma 2.2(c) below, p is a w-polynomial w w mean of φt and P Ew (J, X) is translation invariant. Lemma 2.2. (a) If p ∈ Pw (J, X) and x∗ ◦ p ∈ Ew,0 (J, C) for each x∗ ∈ X ∗ then p ∈ Ew,0 (J, X). (b) On Pw (J, X) the norms kφkw,∞ and |||φ||| = supx∗ ∈X ∗ are equivalent.
||x∗ ◦φ||w,∞ ||x∗ ||
(c) If φ ∈ BCw (J, X) then ∆t φ ∈ Ew,0 (J, X) for all t ∈ J. (d) If φ ∈ P Ew (R, X) has w-mean p, then φ|J ∈ P Ew (J, X) and φ|J has w-mean p|J . (e) Pw (J, X) ⊆ BU Cw (J, X). (f) Let φ ∈ BU Cw (R, X), f ∈ L1w (R) and suppose φ|J is w-polynomially ergodic with w-mean p|J where p ∈ Pw (R, X). Then (φ ∗ f )|J is w-polynomially ergodic with w-mean (p ∗ f )|J . Proof. (a) We can choose q1 , ..., qm ∈ Pw (J, C) Pand linearly independent unit vectors x1 , ..., xm ∈ X such that p = m j=1 qj ⊗ xj . Also choose
∗ ∗ ∗ unit vectors xj ∈ X such that xj , xi = δi,j . Given ε > 0 there
∗
are finite subsets Fj of J such that RF j (xj ◦ p/w) < ε/m. Setting F = F1 + ... + Fm we find
m m
X
X
kRF (p/w)k = RF (qj /w) ⊗ xj ≤ kRF (qj /w)k
=
j=1 m X
j=1 m X
j=1
j=1
RF (x∗j ◦ p/w) ≤
RF (x∗j ◦ p/w) < ε. j
This proves that p ∈ Ew,0 (J, X) (b) Let {p1 , ..., pk } be a basis of Pw (J, C) consisting of unit vecPk tors for the norm kpkw,∞ . If p = j=1 cj pj , where cj ∈ C then Pk kpkw,∞ ∼ j=1 kcj k, ∼ denoting equivalence of norms. Every φ ∈ P Pw (J, X) has a unique representation φ = kj=1 pj ⊗ xj , where xj ∈ P X, and by the closed graph theorem, kφkw,∞ ∼ kj=1 kxj k. Hence,
P Pk
|||φ||| = sup kj=1 pj hx∗ , xj i / kx∗ k ≤ j=1 kxj k ∼ kφkw,∞ . Conversely, choose j0 such that kxj k ≤ kxj0 k for each j. Then choose x∗ ∈ X ∗ such that hx∗ , xj0 i = kxj0 k and kx∗ k = 1. Hence,
ALMOST PERIODIC BEHAVIOUR OF UNBOUNDED SOLUTIONS
21
k
X
|||φ||| ≥ pj hx∗ , xj i
j=1
∼
k X j=1
w,∞ k
|hx∗ , xj i| ≥ kxj0 k ≥
1X kxj k ∼ kφkw,∞ . k j=1
(c) We have ∆wt φ = ∆t ( wφ ) + ( wφ )t ∆wt w where ∆t ( wφ ) ∈ E0 (J, X) by [11, Proposition 3.2] and ( wφ )t ∆wt w ∈ C0 (J, X). (d) We prove the case J = R Given ε > 0 there is a finite subset
+ . P
1 m φ−p
F = {t1 , ..., tm } ⊆ R such that m j=1 ( w )(tj + t) < ε for all t ∈ R. choose uj , vj ∈ R+ such that tj = uj − vj . Let v = v1 + ... + vm and set
P φ−p sj = tj + v. So sj ∈ R+ and m1 m j=1 ( w )(sj + t) < ε for all t ∈ R+ . (e) Given p ∈ P n (J, X) we may choose pj ∈ P n (J, X) and qj ∈ Pk P n (J, C) with qj (0) = 0 such that ∆h p(t) = j=1 pj (t)qj (h) for all Pk h, t ∈ J. Hence k∆h p(t)k ≤ cw(t) j=1 |qj (h)|, where c = supj kpj kw,∞ , and so p ∈ BU Cw (R, X). (f) If χ is the characteristic function of a compact set K ⊆ R then R (φ − p) ∗ χ(s) = −K (φ − p)t (s)dt for each s ∈ R. But for each t ∈ R, (φ − p)t = ∆t (φ − p) + (φ − p) and so by (c), (φ − p)t |J ∈ Ew,0 (J, X). Also, by (e), φ − p ∈ BU Cw (R, X). By (1.1), the function t → (φ − p)t |J : R → Ew,0 (J, X) is continuous andRhence weakly measurable and separably-valued on −K. The integral −K (φ − p)t |J dt is therefore a convergent Haar-Bochner integral and so belongs to Ew,0 (J, X). As evaluation at s ∈ J is continuous on Ew,0 (J, X) we conclude that ((φ − p) ∗ χ)|J ∈ Ew,0 (J, X). Hence also ((φ − p) ∗ σ)|J ∈ Ew,0 (J, X) for any step function σ : R → C By [28, p. 83] the step functions are dense in L1w (R) and so ((φ − p) ∗ f )|J ∈ Ew,0 (J, X) for any f ∈ L1w (R). The difference theorem below, included here in order to characterize Ew (J, X) will also be used later. We use the notation Cw,0 (J, X) = {wξ : ξ ∈ C0 (J, X)}, clearly a closed subspace of BU Cw (J, X). Theorem 2.3. Let F be any translation invariant closed subspace of BCw (J, X). If φ ∈ P Ew (J, X) has w-mean p and ∆t φ ∈ F for each t ∈ J, then φ − p ∈ F + Cw,0 (J, X). If also w = 1, then φ − p ∈ F. P φ−p φ−p 1 Proof. For any finite subset F ⊆ J, φ−p −R ( ) = − F t∈F ∆t ( w ) w w |F | P P and so φ − p = wRF ( φ−p ) − |F1 | t∈F ∆t φ + |F1 | t∈F ( φ−p )∆w+ w w t t P 1 t∈F ∆t p. The first term on the right may be made arbitrarily |F |
22
BOLIS BASIT AND A. J. PRYDE
small in norm by suitable choice of F . The second term is in F by assumption, the third and fourth terms are in Cw,0 (J, X) since ∆t p, ∆t w ∈ Cw,0 (J, X) for t ∈ J. If w = 1 then ∆t w = ∆t p = 0 which shows φ − p ∈ F. We are now able to characterize w-polynomially ergodic functions. Denote by Dw,0 (J, X) the closed span of {∆t φ : t ∈ J, φ ∈ BCw (J, X)}∪ Cw,0 (J, X) and by Dw (J, X) the closed span of {∆t φ : t ∈ J, φ ∈ BCw (J, X)}. Corollary 2.4. Ew,0 (J, X) = Dw,0 (J, X). If w = 1, then Dw,0 (J, X) = Dw (J, X) = E0 (J, X). Proof. Since Cw,0 (J, X) ⊂ Ew,0 (J, X), by Lemma 2.2 (c) and the closedness of Ew,0 (J, X), we have Dw,0 (J, X) ⊂ Ew,0 (J, X). Conversely, let φ ∈ Ew,0 (J, X). Then ∆t φ ∈ Dw (J, X) for all t ∈ J and by Theorem 2.3, φ ∈ Dw,0 (J, X). If w = 1, then for P any ψ ∈ Cw,0 (J, X) and any finite subset F of J, we have ψ = − |F1 | t∈F ∆t ψ +RF ψ. As kRF ψkw,∞ may be made arbitrarily small, we conclude that ψ ∈ Dw (J, X) and hence Dw,0 (J, X) = Dw (J, X). We conclude this section with a characterization of APw (R, X). Theorem 2.5. If N ≥ 1, then APw (R, X) = tN AP (R, X)⊕Cw,0 (R, X). Proof. Note firstly that the sum on the right is direct. For suppose tN ψ1 + ξ1 = tN ψ2 + ξ2 for ψj ∈ AP (R, X) and ξj ∈ Cw,0 (R, X). Set q(t) = (1 + t)N − tN and J = R+ . Then (ψ1 − ψ2 )|J = w1 (ξ2 − ξ1 − qψ2 + qψ1 )|J ∈ C0 (J, X). This is impossible unless ψ1 = ψ2 (see [31] or [5, Proposition 2.1.6). The sum is also topological. For suppose φn = tN ψn + ξn where ψn ∈ AP (R, X) and ξn ∈ Cw,0 (R, X) and n (φn ) converges to φ in BCw (R, X). Then φwn |J = ψn |J + ξn −qψ |J ∈ w AP (R, X)|J ⊕ C0 (J, X). But this last sum is a topological direct sum (see [5, 31]) and so (ψn |J ) converges to ψ|J for some ψ ∈ AP (R, X). Hence (ψn ) converges to ψ in AP (R, X) and tN ψn converges to tN ψ in APw (R, X). It follows that (ξn ) converges to some ξ in Cw,0 (R, X) and that φ = tN ψ + ξ. Next, given φ ∈ APw (R, X) we may choose a sequence (πn ) ⊂ T Pw (R, X) converging to φ in APw (R, X). But πn = tN ψn + ξn where ψn ∈ T P (R, X) and ξn ∈ Cw,0 (R, X). It follows from the previous paragraph that φ = tN ψ + ξ for some ψ ∈ AP (R, X) and ξ ∈ Cw,0 (R, X). Conversely, let φ = tN ψ + ξ for some ψ ∈ AP (R, X) and ξ ∈ Cw,0 (R, X). We may choose αn ∈ T P (R, X) such that kψ(t) − αn (t)k ≤ 1 (see[1, (1.2), p. 15] or [23]). Moreover, wξ ∈ C0 (R, X) and kξ(s)k ≤ n
ALMOST PERIODIC BEHAVIOUR OF UNBOUNDED SOLUTIONS
23
kξk∞,w w(s) for all s. Hence we may choose e tn > 0 such that kξ(t)k ≤ 1 w(t) for all |t| ≥ e tn . Then choose tn > e tn such that kξk∞,w w(e tn ) ≤ n 1 1 1 w(tn ). It follows that kξ(t)k ≤ n w(t) and kξ(s)k ≤ n w(tn ) for n all |t| ≥ tn and all |s| ≤ tn . Since ξ is continuous we may choose βn ∈ T P (R, X) such that kξ(s) − βn (s)k ≤ n1 and kβn (t)k ≤ n1 w(tn )+ n1 for all |s| ≤ tn and all t. Thus kξ(s) − βn (s)k ≤ n3 w(s) for all s. Set πn = tN αn + βn ∈ T Pw (R, X). Then (πn ) converges to φ in BCw (R, X) and so φ ∈ APw (R, X). Corollary 2.6. APw (R, X) ⊂ P Ew (R, X). Proof. Let φ = tN ψ + ξ where ψ ∈ AP (R, X) and ξ ∈ Cw,0 (R, X). Set p = tN M ψ and ψ0 = ψ − M ψ. Then φ−p = ψ0 − ψw0 + wξ on R+ and w φ−p φ−p ψ0 ξ = −ψ0 + w + w on R− . Hence w |J ∈ E0 (J, X) for J = R+ or w ∈ E0 (R, X). R− and thus φ−p w 3. Spectral analysis Throughout this section we will assume that F is a BU Cw -invariant closed subspace of BCw (J, X). A subspace F of BCw (J, X) is called BU Cw -invariant (see [12]) if φt |J ∈ F whenever φ ∈ BU Cw (R, X), φ|J ∈ F and t ∈ R. Numerous examples are provided in [12]. b = {γs : γs (t) = eist for s, t ∈ R} The dual group of R is denoted R R∞ and the Fourier transform of f ∈ L1 (R) by fˆ(γs ) = −∞ f (t)γs (−t)dt. Let φ ∈ BCw (R, X). The set Iw (φ) = {f ∈ L1w (R) : φ ∗ f = 0} is a closed ideal of L1w (R) and the Beurling spectrum of φ is defined b : fˆ = 0 for all γ ∈ Iw (φ)}. to be spw (φ) = cosp(Iw (φ)) = {γ ∈ R More generally, following [5, Section 4], the set IF (φ) = {f ∈ L1w (R) : (φ ∗ f )|J ∈ F } is a closed translation invariant subspace of L1w (R) and therefore an ideal. We define the spectrum of φ relative to F, or the reduced Beurling spectrum, to be spF (φ) = cosp (IF (φ)). The following proposition contains some basic properties of these spectra. The proofs are the same as for the Beurling spectrum. See for example [17, p. 988] or [29] also [6], [15], [27]. Proposition 3.1. Let φ, ψ ∈ BCw (R, X). (a) spF (φt ) = spF (φ) for all t ∈ R. (b) spF (φ ∗ f ) ⊆ spF (φ) ∩ supp(fˆ) for all f ∈ L1w (R). (c) spF (φ + ψ) ⊆ spF (φ) ∪ spF (ψ). (d) spF (γφ) = γspF (φ), provided F is invariant under multiplication b by γ ∈ R.
24
BOLIS BASIT AND A. J. PRYDE
(e) If f ∈ L1w (R) and fˆ = 1 on a neighbourhood of spF (φ), then spF (φ ∗ f − φ) = ∅. The following theorem is proved in [12](see also [10], [11]). It gives our motivation for introducing spF (φ). Theorem 3.2. Let φ ∈ BU Cw (R, X). (a) If f ∈ L1w (G) and φ|J ∈ F, then (φ ∗ f )|J ∈ F. (b) spF (φ) = ∅ if and only if φ|J ∈ F. (c) If ∆kt φ|J ∈ F for all t ∈ R and some k ∈ N, then spF (φ) ⊆ {1}. P (d) spF (φ) ⊆ {γ1 , ..., γn } if and only if φ = ψ + nj=1 ηj γj for some ψ, ηj ∈ BU Cw (R, X) with ψ|J ∈ F and ∆t ηj |J ∈ F for each t ∈ RN +1 . 4. Primitives and Derivatives Throughout this section we assume that F is a translation invariant closed subspace of BU Cw (J, X). Examples of such classes are Pw (J, X), Cw,0 (J, X), APw (R, X), Ew,0 (J, X) ∩ BU Cw (J, X) and P Ew (J, X) ∩ BU Cw (J, X). R t We define the primitive P φ of a function φ ∈ BCw (R, X) by P φ(t) = φ(s)ds. 0 Theorem 4.1. (a) If Fw denotes any of BCw (J, X), Cw,0 (J, X), Ew,0 (J, X), Pw (J, X), P Ew (J, X) or APw (R, X) then P maps Fw continuously into Fww1 . (b) If φ ∈ Ew,0 (J, X) then P φ ∈ Cww1 ,0 (J, X). (c) If φ ∈ APw (J, X) has w-mean p. Then P (φ − p) ∈ Cww1 ,0 (R, X). Proof. Take J = R+ , the other cases being proved similarly. If φ ∈ BCw (J, X) and t ∈ J then ||P φ(t)|| ≤ t.||φ||w,∞ w(t). Hence P maps BCw (J, X) continuously into BCww1 (J, X). If also φ ∈ Cw,0 (J, X) then given ε > 0 there exists t0 > 0 suchR that ||φ(t)|| < εw(t) whenever t > t t0 . For these t we have ||P φ(t)|| ≤ 0 0 ||φ(s)|| ds + εw(t)(t − t0 ) and so P maps Cw,0 (J, X) into Cww1 ,0 (J, X). Next, P (∆t φ) = ∆t (P φ) − P φ(t) and since P is continuous it follows from Corollary 2.4 that P maps Ew,0 (J, X) into Eww1 ,0 (J, X). The result for Pw (J, X) is clear and so therefore is the result for P Ew (J, X). For (b) note that ||∆t P φ(s)|| ≤ t.w(t)w(s) ||φ||w,∞ for all s ∈ J. Hence ∆t (P φ) ∈ Cww1 ,0 (J, X). If φ ∈ Ew,0 (J, X), we can apply Theorem 2.3 to P φ to obtain P φ ∈ Cww1 ,0 (J, X). Finally, (c) follows from (b) using Corollary 2.6, and then (a) with Fw = APw (R, X) follows from (c) using Theorem 2.5.
ALMOST PERIODIC BEHAVIOUR OF UNBOUNDED SOLUTIONS
25
Proposition 4.2. (a) If φ ∈ BCw (R, X) and spw (φ) is compact, then φ(j) ∈ BU Cw (R, X) for all j ≥ 0. (b) If φ ∈ F and φ0 is w-uniformly continuous, then φ0 ∈ F. (c) If φ ∈ BCw (J, X) and φ0 is w -uniformly continuous, then φ0 ∈ Ew,0 (J, X) ∩ BU Cw (J, X). (d) If φ, φ0 ∈ BCw (R, X) then spw (φ0 ) ⊆ spw (φ) ⊆ spw (φ0 ) ∪ {1}. (e) If φ, φ0 ∈ BCw (J, X) then φ ∈ BU Cw (J, X). Proof. (a) Choose f ∈ S(R), the Schwartz space of rapidly decreasing functions, such that f has compact support and is 1 on a neighbourhood of spw (φ). Then f (j) ∈ L1w (R) for all j ≥ 0. Moreover, φ = φ ∗ f and so φ(j) = φ ∗ f (j) for all j ≥ 0. Hence φ(j) ∈ BU Cw (R, X). (b) If ψn = n∆1/n φ then ψn ∈ F. Moreover, by the w-uniform continuity of φ0 , given ε > 0 there exists nε such that
Z
1/n
0 0 0 kψn (t) − φ (t)k = n (φ (t + s) − φ (t))ds < εw(t)
0
for all t ∈ J and n > nε . Hence φ0 ∈ F. (c) With the notation used in the proof of (b), ψn ∈ Ew,0 (J, X) ∩ BU Cw (J, X) by Lemma 2.2(c). Hence, so does φ0 . (d) For any f ∈ L1w (R) we have (φ∗f )0 = φ0 ∗f and so Iw (φ0 ) ⊇ Iw (φ). Hence, spw (φ0 ) ⊆ spw (φ). For the second inclusion, let g(t) = exp(−t2 ) b (spw (φ0 ) ∪ so that g, g 0 ∈ L1w (R) and gˆ is never zero. Now take γ ∈ R {1}). So γ(t) = eist for some s 6= 0 and there exists f ∈ L1w (R) such that φ0 ∗ f = 0 but fˆ(γ) 6= 0. Let h = f ∗ g 0 ∈ L1w (R). Then ˆ φ ∗ h = φ ∗ f ∗ g 0 = φ0 ∗ f ∗ g = 0 whereas h(γ) = isfˆ(γ)ˆ g (γ) 6= 0. So 0 0 γ∈ / spw (φ ) showing spw (φ) ⊆ spw (φ ) ∪ {1}.
R t+h
(e) For any h, t ∈ J we have k∆h φ(t)k = t φ0 (s)ds ≤ |h| · kφ0 kw,∞ w(h)w(t) from which it follows that φ is w-uniformly continuous. Proposition 4.3. Let φ ∈ F and assume that F is BU Cw -invariant. (a) If P φ is w-polynomially ergodic with w-mean p, then P φ − p ∈ F + Cw,0 (J, X). (b) If F = APw (R, X) and P φ ∈ P Ew (R, X), then P φ ∈ APw (R, X). (c) If F = APw (R, X) and P φ ∈ BU Cw (R, X), then spF (P φ) ⊆ {1}. (d) If F = C0 (R+ , X) and P φ is ergodic with mean c, then P φ−c ∈ C0 (R+ , X).
26
BOLIS BASIT AND A. J. PRYDE
Proof. (a) Take J = R+ , the other cases being proved similarly. Exe tend φ to an even function φ ∈ BU CRw (R, X). For t ≥ 0 set χt = χ[−t,0] so that ∆t P φ = φe ∗ χt |J = R φe−s |J χt (s)ds. Since φe ∈ BU Cw (R, X) the integral converges as a Lebesgue-Bochner integral. Since F is BU Cw -invariant φe−s |J ∈ F and therefore ∆t P φ ∈ F. The result follows from Theorem 2.3. (b) In view of Theorem 2.5, this follows from (a). (c) Let s, t ∈ R With χs as in the previous proof, (∆s P φ)t = φt ∗ χs and by Proposition 3.2 (a), (∆s P φ)t ∈ F. By Proposition 3.2(c), spF (P φ) ⊆ {1}. (d) This is a special case of part (a). Remark 4.4. Recall u(t) = 12 w(t) cos log w(t) + 21 w(t) sin log w(t) − t sin log w(t) − 12 from Example 1.1. So u0 (t) = w(t) cos log w(t) and 0 therefore u ∈ Cw,0 (R, C) ⊂ APw (R, C) ⊂ P Ew (R, C). However, u ∈ / P Ew (R, C). Indeed, if u ∈ P Ew (R, C) then for t ∈ R+ set ξ(t) = w(t) cos log w(t) + w(t) sin log w(t). So ξ ∈ P Ew (R+ , C) and for some polynomial p(t) = at + b we have (ξ − p)/w ∈ E0 (R+ , C). Thus η = ξ/w ∈ E(R+ , C). But η 0 (t) = [− sin log w(t) + cos log w(t)]/w(t) and so η 0 ∈ C0 (R, C). By Proposition 4.3(d) we conclude η ∈ C0 (R+ , C) + C which is false. Lemma 4.5. For natural numbers m, N and non-negative integers j, k set a(m, j) = (−1)j Nj m−1+j j!. j P N −j m+j (a) P m (tN φ) = N P φ for any φ ∈ L1loc (J, X). j=0 a(m, j) t N +k−m N ! if m ≤ k N N (N +k)! X a(m, j) = 0 (b) if k + 1 ≤ m ≤ k + N . N! (j + k)! m−k−1 N j=0 (−1) if m > k + N N (N +k)! (c)
PN
a(m, j) tN −j P j+1 r = P r ∈ Pm−2 (J, X). j=0
PN
j=0
a(m − 1, j) tN −j P j r for any
Proof. (a) For N = 1 the claim is readily proved by induction on m. The general case is then proved by induction on N . (b) For m ≥ k + 1 we have PN
a(m,j) j=0 (j+k)!
N X
N m−1+j j! = (−1) j j (j + k)! j=0 j
ALMOST PERIODIC BEHAVIOUR OF UNBOUNDED SOLUTIONS
27
N X 1 (m − 1 + j)! j N = (−1) (m − 1)! j=0 j (k + j)! N X 1 j N (−1) Dm−k−1 tm+j−1 |t=1 = (m − 1)! j=0 j N X 1 m−k−1 m−1 j N = D t (−1) tj |t=1 (m − 1)! j j=0 1 Dm−k−1 tm−1 (1 − t)N |t=1 . (m − 1)! For m − k − 1 < N this last expression is 0 and for m − k − 1 ≥ N it is 1 m−k−1 (Dm−k−1−N tm−1 )DN (1 − t)N |t=1 (m − 1)! N 1 m − k − 1 (m − 1)! (−1)N N ! = (m − 1)! N (N + k)! as claimed. For m ≤ k the claim follows readily by substituting φ(t) = tk−m in (a). P a(m−1,j) (c) It follows readily from (b) that N j=0 (j+k)! = (N + k + 1) × PN Pm−2 k a(m,j) j=0 (j+k+1)! if 0 ≤ k ≤ m − 2. So setting r(t) = k=0 ck t we find =
PN
j=0
a(m, j) tN −j P j+1 r(t) =
m−2 X
N +k+1
ck k!t
j=0
k=0
=
m−2 X
ck k!t
m−2 X
ck k!tN +k
k=0
=P
a(m, j) (j + k + 1)! N
N +k+1
k=0
=P
N X
N X
X a(m − 1, j) 1 N + k + 1 j=0 (j + k)! N X a(m − 1, j) (j + k)! j=0
a(m − 1, j) tN −j P j r(t).
j=0
Our main result is the following: Theorem 4.6. Assume φ ∈ AP (R, X) and that BU CwN (R, X) for some bj ∈ C, b0 6= 0.
PN
j=0 bj
tN −j P j+1 φ ∈
28
BOLIS BASIT AND A. J. PRYDE
(a) P φ ∈ BU C(R, X) and if
bj j=0 (j+1)!
PN
6= 0 then M φ = 0.
(b) If X 6⊇ c0 then P (φ − M φ) ∈ AP (R, X). P N −j j+1 Proof. Let a = M φ and ψ = N P φ. Then we have ψ = j=0 bj t PN P b N j bj tN −j P j+1 (φ − a) + tN +1 a j=0 (j+1)! . By Theorem 4.1(c), ψ − Pj=0 N N −j j+1 P (φ − a) ∈ CwN +1 ,0 (R, X) and so either a = 0 or j=0 bj t PN bj j=0 (j+1)! = 0. To prove the rest of the theorem, we may assume a = 0. By Theorem 4.1(c), P j φ(t)/wj (t) → 0 as t → ∞. Since φ is almost periodic we may choose (tn ) ⊂ R such that tn → ∞ and φtn → φ uniformly on R. Moreover, as M φ = 0, by Theorem 4.1(c), P j φ(s+tn )/wj (s+tn ) → 0 uniformly on R for j > 0. Given x∗ ∈ X ∗ , it follows that x∗ ◦P φtn → x∗ ◦P φ locally uniformly. Moreover, by passing to a subsequence if necessary, we may assume x∗ ◦ ψ(tn )/wN (tn ) → b for some b ∈ C. By Theorem 4.1(c) again, we obtain Z t N X N −j ψ(t + tn ) = bj (t + tn ) [ P j φ(s + tn )ds + P j+1 φ(tn )] j=0
0
N = ψ(tn ) + b0 tN n P φ(t + tn ) + o(tn ).
Therefore x∗ ◦ ψ(t + tn )/wN (t + tn ) → b + b0 x∗ ◦ P φ(t) for each t ∈ R. Hence, since ψ/wN is bounded, so too is x∗ ◦ P φ. Since x∗ is arbitrary, P φ is weakly bounded and therefore bounded. From Proposition 4.2(e) it follows that P φ ∈ BU C(R, X). If also X 6⊇ c0 then by Kadet’s theorem [21] (see also [4]), P φ is almost periodic. Corollary 4.7. Assume φ ∈ AP (R, X) and P (tN φ) ∈ BU CwN (R, X). (a) P φ ∈ BU C(R, X) and M φ = 0. (b) If X 6⊇ c0 then P φ ∈ AP (R, X). PN N −j m+j Proof. Since P m (tN φ) = P φ the result follows j=0 a(m, j) t from Theorem 4.6 and Lemma 4.5. Theorem 4.8. Assume φ ∈ AP (R, X), X 6⊇ c0 , P m (tN φ) + p ∈ BU CwN (R, X) for natural numbers m, N and some p ∈ Pm−1 (R, X). Then P j (tN φ) + p(m−j) ∈ APwN (R, X) for 1 ≤ j ≤ m. Moreover, there is a polynomial q ∈ Pm−1 (R, X) such that P j φ + q (m−j) ∈ AP (R, X) P P N −j j k P q= for 1 ≤ j ≤ m and, if p(t) = m−1 then N j=0 a(m, j) t k=0 bk t Pm−1 k k=N +1 bk t . Proof. The proof is by induction on m. If m = 1 then, by Lemma PN a(1,j) PN 1 N −j j+1 4.5 and P (tN φ) = P φ ∈ j=0 (j+1)! = N +1 j=0 a(1, j) t
ALMOST PERIODIC BEHAVIOUR OF UNBOUNDED SOLUTIONS
29
BU CwN (R, X). Therefore, by Theorem 4.6, M φ = 0 and P φ ∈ AP (R, X). Moreover, by Lemma 4.5(b), N X
N −j
a(1, j) t
j
N
P (M P φ) = (M P φ) t
j=0
N X aj j=0
j!
= 0.
PN
N −j j Hence P (tN φ) = P (P φ − M P φ) and by Theorem j=0 a(1, j) t N 4.1(c), P (t φ) ∈ APwN (R, X). For m > 1, Theorem 5.2 below shows P j (tN φ) + p(m−j) ∈ BU CwN (R, X) for 1 ≤ j ≤ m. Hence, as induction hypothesis we may assume there is a polynomial r ∈ Pm−2 (R, X) such that for 1 ≤ j ≤ m−1 we have P j φ+r(m−1−j) ∈ AP (R, X), P j (tN φ)+ P P N −j j k−1 p(m−j) ∈ APwN (R, X) and N P r = m−1 . j=0 a(m−1, j) t k=N +2 kbk t m−1 In particular, η = P φ + r + c ∈ AP (R, X) where the constant c is P N −j j+1 to be chosen. Moreover, by Lemma 4.5(d), N P r= j=0 a(m, j) t Pm−1 k k=N +2 bk t . Now set q = P (r + c − M φ) so that P m φ + q = P (η − M φ). By Theorem 4.6, to show P m φ + q ∈ AP (R, X), it suffices to show that PN N −j j+1 P η ∈ BU CwN (R, X). By Lemma 4.5(a), j=0 a(m, j) t m
N
P (t φ) =
N X
a(m, j) tN −j P m+j φ
j=0
=
N X
a(m, j) tN −j P j+1 (η − r − c).
j=0
P Since P m (tN φ)+p ∈ BU CwN (R, X), it suffices to show N j=0 {a(m, j)× Pm−1 k N −j j+1 t P (r + c) = k=N +1 bk t . If N > m − 2 we choose c = 0 as then both sides are 0. Otherwise N ≤ m − 2 and by Lemma 4.5(c) we may choose c such that PN P a(m,j) N −j j+1 P c = bN +1 tN +1 , that is c N j=0 a(m, j) t j=0 (j+1)! = bN +1 . In we have Theorem 4.6, M η = 0. InPeither case PNthis case also Pby N N −j j k P q = j=0 a(m, j) tN −j P j+1 (r + c) = m−1 j=0 a(m, j) t k=N +1 bk t PN N −j j+1 and P M η = 0. Finally, j=0 a(m, j) t P m (tN φ) + p =
N X
a(m, j) tN −j P j+1 (η − r − c − M φ) + p
j=0
=
N X
N −j
a(m, j) t
j
m
P (P φ + q) +
j=0 m
N X
bk tk
k=0 N
and by Theorem 4.1, P (t φ) + p ∈ APwN (R, X).
30
BOLIS BASIT AND A. J. PRYDE
Remark 4.9. (a) In Theorem 4.6(a) the space AP (R, X) may be replaced by the class of Poisson stable functions. These are functions ξ ∈ C(R, X) for which there exist sequences (tn ) ⊂ R such that tn → ∞ and ξtn → ξ locally uniformly on R. In part (b), AP (R, X) may be replaced by any class for which Kadet’s theorem remains valid. These include Poisson stable functions, almost automorphic functions and recurrent functions (see [4 ]). P N −j j (b) If p = 0 in Theorem 4.8 then N P q = 0, which j=0 a(m, j) t (k) reduces to q (0) = 0 for 0 ≤ k ≤ m − N − 1. (c) Assume φ ∈ AP (R, X) where X 6⊇ c0 . By Theorem 4.8, if m N P (t φ) + p ∈ BU CwN (R, X) for some p then P m φ + q ∈ AP (R, X) for some q. PN The converse is also true. Indeed, P m (tN φ) = j=1 {a(m, j)× N −j m+j N m t P φ + t P φ and the result follows from Theorem 4.1(c). (d) These results are dependent on the Poisson stability property of 1 φ. Indeed, consider the function φ ∈ C0 (R, C) given by φ(t) = 1+|t| . Then P (tφ) = |t| − ln(1 + |t|) and P φ = sgn(t) ln(1 + |t|). Hence P (tφ) ∈ BU Cw1 (R, C) whereas P φ ∈ / BC(R, C). (e) A well-known example to show that the condition X 6⊇ c0 may not be omitted from Theorem 4.6 is as follows. Let X = c0 and φ(t) = 2 t ∞ ( n1 sin nt )∞ n=1 so that P φ(t) = (2 sin 2n )n=1 . Then φ ∈ AP (R, c0 ) and P φ ∈ BU C(R, c0 ). However, P φ does not have relatively compact range so it is not almost periodic. 5. Esclangon-Landau theorem In this section we use the abbreviations m X (5.2) Bu = bj u(j) j=0
and assume bm = 1, bj ∈ C, u : J → X. We prove a theorem of Esclangon-Landau type ([18], [22], [14], [7], [16] and references therein). Lemma 5.1. If Bu = ψ where u, ψ ∈ BCwN (J, X) then u(j) (t) = O(|t|N +m−1 ) for 1 ≤ j ≤ m. P (j) m−k Proof. Since u(m) = ψ − m−1 we obtain j=0 bj u , taking P u
(k)
=
m−k X j=1
P
j−1
(u
(j+k−1)
(0)) + P
m−k
ψ−
m−1 X j=0
bj P m−k u(j) .
ALMOST PERIODIC BEHAVIOUR OF UNBOUNDED SOLUTIONS
31
Setting k = 1 we conclude that u0 (t) = O(|t|N +m−1 ). In general P m−k (j) u(k) (t) = O(|t|N +m−1 ) − m−1 u (t) = O(|t|N +m−1 ) + j=m−k+1 bj P Pk−1 (j) O( u (t) ) from which the result follows by induction. j=1
Theorem 5.2. If Bu = ψ where u, ψ ∈ BCwN (J, X) then u(j) ∈ BCwN (J, X) for 1 ≤ j ≤ m. Proof. Take J = R+ , the other cases being proved similarly. The proof is by induction on m. First, if m = 1 the equation becomes u0 + b0 u = ψ showing u0 ∈ BCwN (J, X). For the general case we use the functions f and u˜ defined by f (t) = exp(−t) for t ≥ 0, f (t) = 0 for t < 0, u˜(t)R = u(−t) for −tR ∈ J and u˜(t) = 0 for −t ∈ / J. ∞ −s t ∞ −s It follows that e t e u(s)ds = 0 e u(s + t)ds = u˜ ∗ f (−t) and u˜ ∗ f ∈ BCwN (R, X). Moreover, using repeated integration by parts R P t ∞ −s (k) (j) and Lemma 5.1, we find e t e u (s)ds = − k−1 ˜∗f j=1 u (t) + u (−t). Hence the equation Bφ = ψ may be transformed to the equation Pm Pk−1 (j) Pm u ∗ f (−t) − ψ˜ ∗ f (−t). This is an k=1 bk j=1 u (t) = ( k=0 bk )˜ (j) equation of order m − 1 and so by the induction hypothesis Pm−1 u (j) ∈ (m) BCwN (J, X) for 1 ≤ j ≤ m − 1. Hence u = ψ − j=0 bj u ∈ BCwN (J, X) which finishes the proof. 6. Application P (j) and assume bm = Again we use the abbreviation Bu = m j=0 bj u 1. By pB we denote the characteristic polynomial of the differential Pm j operator B. Thus pB (s) = j=0 bj (is) and for smooth f we have c (γs ) = pB (s)fˆ(γs ). The set of complex zeros of pB is denoted Z(B). Bf Lemma 6.1. Assume u ∈ BCwN (R, X) and F is a BU CwN -invariant closed subspace of BCwN (J, X). If Bu = ψ where ψ ∈ BU CwN (R, X) and ψ|J ∈ F then spF (u) ⊂ {γs : s ∈ Z(B) ∩ R}. Proof. Take s ∈ R with pB (s) 6= 0. Choose f ∈ S(R) with fˆ(γs ) 6= 0 and set g = Bf . Then u∗g = ψ ∗f and by Theorem 3.2(a), (ψ ∗ f ) |J ∈ F. Hence g ∈ IF (u) whereas gˆ(γs ) = pB (s)fˆ(γs ) 6= 0. So γs ∈ / spF (u) and the proof is completed. Theorem 6.2. Suppose Bu = ψ where u ∈ BCwN (R, X) and ψ ∈ APwN (R, X). (a) If Z(B) ∩ R = ∅ then u(j) ∈ APwN (R, X) for 0 ≤ j ≤ m. (b) If Z(B) ∩ R 6= ∅, but X 6⊇ c0 and ψ = tN φ where φ ∈ AP (R, X) then u(j) ∈ APwN (R, X) for 0 ≤ j ≤ m.
32
BOLIS BASIT AND A. J. PRYDE
Proof. (a) Let F = APwN (R, X). By Lemma 6.1, spF (u) ⊂ Z(B)∩R = ∅. Hence, by Theorem 3.2(b), u ∈ F. The Esclangon-Landau Theorem 5.2 shows u, u0 , ..., u(m) ∈ BCwN (R, X) and then Proposition 4.2(e) shows u, u0 , ..., u(m−1) ∈ BU CwN (R, X). From Proposition 4.2(b) we conclude u0 , ..., u(m−1) ∈ F. Rearranging the differential equation, we obtain u(m) ∈ F. (b) The proof is by induction on m. Note first that by Theorem 5.2, u(j) ∈ BCwN (R, X) for 0 ≤ j ≤ m. Let λ ∈ Z(B) ∩ R and make the substitution η(t) = exp(−iλt)u(t) so that η (j) ∈ BCwN (R, X) for 0 ≤ j ≤ m. If m = 1 the equation u0 − iλu = tN φ reduces to η 0 = exp(−iλt)tN φ. From Theorem 4.8 we conclude η ∈ F. Hence u, u0 ∈ F as claimed. For general equation Bu = tN φ reduces to Pmm, the (j) an equation of the form j=1 cj η = exp(−iλt)tN φ where cm = 1. This is a differential equation in η 0 of order m − 1. By the induction hypothesis, or by part (a) if the characteristic polynomial has no real zeros, η (j) ∈ F for 1 ≤ j ≤ m − 1. It to show η ∈ F. For Premains m (j) this, let k = min{j : cj 6= 0}. From j=k cj η = exp(−iλt)tN φ we P (j−k) = P k (exp(−iλt)tN φ) + p for some polynomial p of obtain m j=k cj η degree at most k − 1. But η (j) ∈ BCwN (R, X) for 0 ≤ j ≤ m and so by Theorem 4.8 again we conclude P k (exp(−iλt)tN φ) ∈ F. Since ck 6= 0 we can rearrange the differential equation and obtain η ∈ F. Remark 6.3. The asymptotic behaviour of bounded solutions of equations more general than (5.1) are investigated by numerous authors (see [2], [3], [6], [8], [13], [26], [27], [30]). In particular, it follows from [12, Theorem 4.7] that if φ ∈ BU Cw (J, X), spAPw (φ) is countable and γ −1 φ ∈ Ew (J, X) for all γ ∈ spAPw (φ), then φ ∈ APw (R, X). In this paper for solutions of (5.1) we have replaced the ergodicity condition by X 6⊇ c0 . This is satisfied, in particular, if X is finite dimensional or reflexive or weakly sequentially complete. So, the results of Theorems 4.6, 6.2 are new even for X = R or C. References [1] L. Amerio and G. Prouse, Almost-Periodic Functions and Functional Equations, Van Nostrand, 1971 [2] W. Arendt and C.J.K. Batty, Almost periodic solutions of first and second order Cauchy problems, J. Differential Equations 137 (1997), 363-383 [3] W. Arendt and C.J.K. Batty, Asymptotically almost periodic solutions of of the inhomogeneous Cauchy problems on the half line, London Math. Soc. 31(1999), 291-304 [4] B. Basit, Generalization of two theorems of M.I.Kadets concerning the indefinite integral of abstract almost periodic functions, Math. Notes 9 (1971), 181-186
ALMOST PERIODIC BEHAVIOUR OF UNBOUNDED SOLUTIONS
33
[5] B. Basit, Some problems concerning different types of vector valued almost periodic functions, Dissertationes Math. 338 (1995), 26 pages [6] B. Basit, Harmonic analysis and asymptotic behavior of solutions to the abstract Cauchy problem, Semigroup Forum 54 (1997), 58-74 [7] B. Basit and H. G¨ unzler, Generalized Esclangon-Landau Conditions for Differential-Difference Equations, J. Math. Analysis and applications 221 (1998), 595-624 [8] B. Basit and H. G¨ unzler, Asymptotic behavior of solutions of neutral equations, J. Differential Equations 149 (1998), 115-142 [9] B. Basit and H. G¨ unzler, Generalized vector valued almost periodic and ergodic distributions, (Monash University, Analysis Paper 113, September 2002, 65 pages), submitted [10] B. Basit and A.J. Pryde, Polynomials and functions with finite spectra on locally compact abelian groups, Bull. Austral. Math. Soc. 51 (1995), 33-42 [11] B. Basit and A.J.Pryde, Ergodicity and differences of functions on semigroups, J. Austral. Math. Soc. 149 (1998), 253-265 [12] B. Basit and A.J.Pryde, Ergodicity and stability of orbits of unbounded semigroup representations, submitted to J. Austral. Math. Soc. [13] C.J.K. Batty, J. V. Neerven and F. R¨abiger, Tauberian theorems and stability of solutions of the Cauchy problem, Trans. Amer. Math. Soc. 350 (1998), 2087-2103 ¨ [14] H. Bohr and O. Neugebauer, Uber lineare Differentialgleichungen mit konstanten Koeffizienten und fastperiodischer rechter Seite, Nachr. Ges. Wiss. G¨ ottingen, Math.-Phys. Klasse (1926), 8-22 [15] Y. Domar, Harmonic analysis based on certain commutative Banach algebras, Acta Math. 96 (1956), 1-66 [16] R. Doss, On the almost periodic solutions of a class of integro-differentialdifference equations, Ann. Math. 81 (1965), 117-123 [17] N. Dunford and J. T. Schwartz, Linear Operators, Part II: Spectral Theory, Interscience Pub., New York, London, 1967 [18] E. Esclangon, Nouvelles reserches sur les fonction quasi p´eriodiques, Ann. Observatoire Bordeaux 16 (1921), 51-177 [19] K. Iseki, Vector valued functions on semigroups I, II, III. Proc. Japan Acad. 31 (1955), 16-19, 152-155, 699-701 [20] K. Jacobs, Ergodentheorie und Fastperiodische Funktionen auf Halbgruppen, Math. Zeit. 64 (1956), 298-338 [21] M. I. Kadets, On the integration of almost periodic functions with values in Banach spaces, Functional Analysis Appl. 3 (1969), 228-230 ¨ [22] E. Landau, Uber einen Satz von Herrn Esclangon, Math. Ann. 102 (1930), 177-188 [23] B.M. Levitan and V.V. Zhikov, Almost Periodic Functions and Differential Equations, Cambridge Univ. Press, Cambridge, 1982 [24] L.H. Loomis, Spectral characterization of almost periodic functions, Ann. of Math. 72 (1960), 362-368 [25] W. Maak, Integralmittelwerte von Funktionen auf Gruppen und Halbgruppen, J. reine angew. Math. 190 (1952), 40-48 [26] V.Q. Ph´ ong, Semigroups with nonquasianalytic growth, Studia Mathematica 104 (1993), 229-241
34
BOLIS BASIT AND A. J. PRYDE
[27] J. Pr¨ uss, Evolutionary Integral Equations and Applications, Birkh¨auser Verlag, Basel, Boston, Berlin, 1993 [28] H. Reiter, Classical Harmonic Analysis and Locally Compact Groups, Oxford Math. Monographs, Oxford Univ., 1968 [29] W. Rudin, Harmonic Analysis on Groups, Interscience Pub., New York, London, 1962 [30] W.M. Ruess and V.Q. Ph´ong, Asymptotically almost periodic solutions of evolution equations in Banach spaces, J. Differential Equations 122 (1995), 282-301. [31] W. M. Ruess and W. H. Summers, Ergodic theorems for semigroups of operators, Proc. Amer. Math. Soc. 114 (1992), 423-432 School of Mathematical Sciences, Monash University, PO Box 28M, VIC 3800, Australia E-mail address:
[email protected] School of Mathematical Sciences, Monash University, PO Box 28M, VIC 3800, Australia E-mail address:
[email protected]
GAUSSIAN UPPER BOUNDS FOR HEAT KERNELS OF A CLASS OF NONDIVERGENCE OPERATORS XUAN THINH DUONG AND EL MAATI OUHABAZ Abstract. Let Ω be a subset of a space of homogeneous type. Let A be the infinitesimal generator of a positive semigroup with Gaussian kernel bounds on L2 (Ω). We then show Gaussian heat kernel bounds for operators of the type bA where b is a bounded, complex valued function.
1. Introduction Behaviour of heat kernels has long been an active topic in functional analysis and partial differential equations. In the past few years, it is known that heat kernel bounds such as Gaussian bounds or Poisson bounds imply various useful properties for operators such as Lp spectral invariance [1], [9], bounded holomorphic functional calculi on Lp spaces [11], Lp − Lq maximal regularity for abstract Cauchy problems [12], [7], Lp -analyticity of the semigroup [15]. A large class of divergence form differential operators on the Euclidean space RD are known to possess Gaussian heat kernel bounds, see [8], [2], [4] and their references. However, nondivergence operators with L∞ coefficients, or even with uniformly continuous coefficients, do not possess Gaussian bounds in general, [5], [6]. Hence we can only hope Gaussian heat kernel bounds for specific classes of nondivergence form operators. The nondivergence form operators −b∆ on the Euclidean space RD were studied in [14]. It was proved that if b : X → C is any bounded measurable function on RD such that
0 for a.e. x ∈ X −tbA
(1)
then the kernel kt (x, y) of the semigroup e has an upper bound with polynomial decay in |x − y|. It was also observed that one can improve it to exponential decay by controlling the constants in the upper bound. The proof in [14] depends on the specific Laplacian ∆ through certain estimates using Sobolev embedding and contraction property of the heat semigroup e−t∆ on RD . Note that since b is complex-valued, the semigroup e−tb∆ is no longer contractive. 35
36
XUAN THINH DUONG AND EL MAATI OUHABAZ
This result was extended in [10] where it was shown that if the heat kernel pt (x, y) of an operator A has a Gaussian bound C
d2 (x,y)
e−c t . (2) tD/2 then a similar upper bound holds for the heat kernels of bA. The proof in [10] relies on the Trotter product formula and estimates on the resolvents. During a seminar, the second named author was asked by E. B. Davies if the result in [10] was still true in more general setting like manifolds where the heat kernel bounds take the form d2 (x,y) C √ e−c t . 0 ≤ pt (x, y) ≤ v(x, t) 0 ≤ pt (x, y) ≤
This is a natural question and this paper is to give a positive answer when the underlying space is a space of homogeneous type or a subset of a space of homogeneous type. Throughout this paper, C, C 0 , c and c0 denote positive constants whose value may change from line to line 2. Main result Let (X, d, µ) denote a metric space equipped with a σ−finite measure µ. We assume that X satisfies the doubling property v(x, 2r) ≤ M v(x, r) ∀x ∈ X, ∀r > 0,
(3)
where M is a constant and v(x, r) denotes the volume of the ball with center x and radius r. Suppose that −A is the generator of a bounded analytic semigroup e−tA on L2 (X, µ) which has a kernel pt (x, y) satisfying a Gaussian upper bound d2 (x,y) C √ e−c t (4) 0 ≤ pt (x, y) ≤ v(x, t) for some positive constants C, c and for all t > 0. The aim of this paper is to show that if b : X → C is any bounded measurable function on X which satisfies condition (1), then e−tbA has a kernel kt (x, y) which satisfies a similar Gaussian upper bound, that is 2 C0 0 d (x,y) √ e−c t |kt (x, y)| ≤ (5) v(x, t) for some positive constants C, c and for all t > 0. As we mentioned above if v is polynomial in r and independent of the centre x, i.e., v(x, r) = crD for all x ∈ X, r > 0, condition √ (4) becomes condition (2) and a similar estimate to (5) (with v(x, t) replaced by
GAUSSIAN UPPER BOUNDS FOR HEAT KERNELS
37
tD/2 ) was shown in [10]. In a slightly more general setting, Gaussian lower bounds of kt (x, y) were also studied in [16]. The proof of our main result relies on the same strategy as in [10] in the sense that it is based on the estimates of powers of the resolvents and the Trotter product formula semigroups. However, the local na√for−1 ture of the upper bound (v(x, t)) requires different approach from the case of the uniform bound t−D/2 . Indeed, an uniform bound on the kernel of an operator can be obtained through the cross norm of the operator itself from the space L1 to the space L∞ but this approach √ −1is clearly not enough to produce the upper bound of the type (v(x, t)) . We overcome this problem by using the resolvent equation which implies the representation (8), together with careful estimates on their kernels; see estimates (9) and (10). We also need to know how to pass from the upper bound (4) on heat kernels to estimates on kernels of powers of the resolvents and vice versa. The following theorem gives that equivalence. Theorem 1. Suppose that −A is the generator of a semigroup e−tA which is bounded analytic with angle ν on L2 (X, µ). The following assertions are equivalent (α) e−tA has a kernel pt (x, y) which satisfies the estimate (5) for all t>0 (β) For all λ > 0 and large enough integer m, (λI + A)−m has a kernel Rλ,m (x, y) which satisfies √ C −c |λ|d(x,y) (6) |Rλ,m (x, y)| ≤ e |λ|m v(x, √1 ) |λ|
(γ) for all θ ∈ [0, ν), λ ∈ Σ(θ + π2 ) and large enough integer m, (λI + A)−m has a kernel Rλ,m (x, y) which satisfies (6) where Σ(α) denotes the sector {z ∈ C, |arg(z)| < α}. Remark. (a) The doubling property (3) implies the strong homogeneity property ∀x ∈ X, r > 0, λ ≥ 1, v(x, λr) ≤ M λn v(x, r) (7) for some n > 0. This property will be used in our proof. (b) The condition m large enough in the above theorem can be replaced by m > n. (c) The main result of this paper is Theorem 2 but Theorem 1 is also of independent interest. We will use Theorem 1 and the Trotter product formula to prove the following.
38
XUAN THINH DUONG AND EL MAATI OUHABAZ
Theorem 2. Assume that −A generates a bounded analytic semigroup which has a kernel pt (x, y) satisfying the upper bound (4). Assume that b ∈ L∞ (X, µ, C), satisfies (1) and the operator −bA generates a bounded analytic semigroup e−tbA on L2 (X, µ). Then e−tbA has a kernel kt (x, y) which satisfies (5). Remarks. (a) Sufficient conditions in terms of b and A which ensure that −bA generates a bounded analytic semigroup are given in [10]. For example, −bA always generates a bounded analytic semigroup if A is self-adjoint. (b) In the above theorem we assumed that pt (x, y) is positive but the result is still true if we replace the assumption (4) by d2 (x,y) C √ e−c t |pt (x, y)| ≤ ht (x, y) ≤ (40 ) v(x, t) where ht (x, y) is a positive heat kernel. In this case, the right hand side of the estimate (14) below will be kb−1 k∞ (λc0 I + H)−1 |f |, where H denotes the generator of the semigroup whose kernel is ht (x, y). The rest of the proof needs no change. As an example, a GaussPwe obtain ∂ 2 ian bound for the heat kernel of bA where A = D ( − ia k ) the k=1 ∂xk magnetic Laplacian on RD . It is well known that the heat kernel of A satisfies (40 ) with ht (x, y) the classical heat kernel of the Laplacian, see [13] for example. We can replace the space X which satisfies the doubling property (3) by Ω where Ω is any subet of X and the result of Theorem 2 is still true. More specifically, the following theorem can be proved by using the same proof as that of Theorem 2. Theorem 3. Assume that −A generates a bounded analytic semigroup on L2 (Ω) which has kernel pt (x, y) satisfying d2 (x,y) C (400 ) 0 ≤ pt (x, y) ≤ X √ e−c t v (x, t) √ X for all t > 0, all x, y ∈ Ω, where v (x, t) denotes the volume of √ the ball with centre x, radius t in the space X. Assume that b ∈ L∞ (Ω, µ, C), satisfies (1) and the operator −bA generates a bounded analytic semigroup e−tbA on L2 (Ω). Then e−tbA has a kernel kt (x, y) which satisfies |kt (x, y)| ≤ for all t > 0, all x, y ∈ Ω.
2 C0 0 d (x,y) √ e−c t v X (x, t)
(500 )
GAUSSIAN UPPER BOUNDS FOR HEAT KERNELS
39
Remarks. In Theorems 2 and 3, we only give heat kernel bounds on kt (x, y) for t > 0. Using Proposition 3.3 of [11], we can extend the results to obtain similar heat kernel bounds on kz (x, y) for complex z in some sector of the complex plane. Applications We give a few applications of our Theorems 2 and 3: (a) Let A be the Laplace-Beltrami operator on a manifold M which satisfies a Sobolev inequality and the doubling property. Then A generates a positive semigroup with Gaussian heat kernel bounds (4). It follows from Theorem 2 that the semigroup e−tbA has Gaussian heat kernel bounds (5). (b) Let A be a divergence form elliptic operator with real, symmetric coefficients acting on a bounded domain Ω of RD with Neumann boundary conditions. Assume that the boundary of Ω satisfies the extension property. Then A generates a positive semigroup with Gaussian heat kernel bounds [8]. More specifically, 1
d2 (x,y)
}e−c t (4000 ). tD/2 It follows from Theorem 3 that the semigroup e−tbA has Gaussian heat kernel bounds 2 1 0 d (x,y) |kt (x, y)| ≤ C 0 max{1, D/2 }e−c t (5000 ). t (c) As a consequence of heat kernel bounds, the two operators bA in applications (a) and (b) above have the following properties: (i) Lp spectral invariance: The connected components of their resolvent sets which contain the positive real line on Lp spaces are independent of p, 1 ≤ p ≤ ∞, (ii) The Lp − Lq maximal regularity for abstract Cauchy problems, (iii) If bA has a bounded holomorphic functional calculus on L2 , then it has a bounded holomorphic functional calculus on Lp , 1 < p < ∞. 0 ≤ pt (x, y) ≤ C max{1,
3. The Proofs. Proof of Theorem 1. We first show (α) ⇒ (β). Let λ > 0. The Laplace transform gives Z ∞ 1 −m (λI + A) = tm−1 e−λt e−tA dt. m! 0
40
XUAN THINH DUONG AND EL MAATI OUHABAZ
Hence the kernel Rλ,m (x, y) of (λI + A)−m is given by Z ∞ m−1 −λt −cd2 (x,y) C t e √ e t dt. |Rλ,m (x, y)| ≤ m! 0 v(x, t) √ 2 ≥ λd(x, y), we have Using the fact that λt + d (x,y) t " √ Z ∞ m−1 −c”λt # Z λ−1 m−1 −c”λt 0 Ce−c λd(x,y) t e t e √ dt + √ dt . |Rλ,m (x, y)| ≤ m! v(x, t) t) λ−1 v(x, 0 √ √ For t ∈ [λ−1 , ∞), we obviously have v(x, t) ≥ v(x, λ−1 ). Hence the second term in the square bracket satifies Z ∞ Z ∞ m−1 −c”λt t e 1 √ √ dt ≤ tm−1 e−c”λt dt −1 t) v(x, λ ) λ−1 λ−1 v(x, Z ∞ 1 √ = λ−m sm−1 e−c”s ds −1 v(x, λ ) 1 C √ = . λm v(x, λ−1 ) For the first term of the square bracket, we have Z λ−1 m−1 −c”λt Z 1 m−1 −c”s 1 t e s e √ dt = m p s ds. λ 0 v(x, λ ) v(x, t) 0 We now apply the strong homogeneity property (7) to deduce that r √ −n s −1 v(x, λ ) ≤ M s 2 v(x, ) ∀s ∈ (0, 1]. λ This implies that Z λ−1 m−1 −c”λt Z 1 n t e C √ dt ≤ √ sm− 2 −1 e−c”s ds v(x, t) λm v(x, λ−1 ) 0 0 and the last integral is finite for m > n2 . This shows the desired implication. We now show (β) ⇒ (γ). Firstly, we show an upper bound for Rm,λ (x, y) for λ ∈ Σ(θ + π2 ). By the resolvent equation ( iterated m−times ) we have (λI + A)−2m = (λI + A)−m (I + (|λ| − λ)(λI + A)−1 )2m (λI + A)−m from which it follows that Z Rλ,2m (x, y) = R|λ|,m (x, z)(LR|λ|,m (., y))(z)dµ(z) X
(8)
GAUSSIAN UPPER BOUNDS FOR HEAT KERNELS
41
where L = (I + (|λ| − λ)(λI + A)−1 )2m . In order to apply L to z → R|λ|,m (z, y) we need to know that the latter is in L2 (X, µ). Let us assume this for the moment. By Cauchy-Schwarz inequality and (8) we have |Rλ,2m (x, y)| ≤ kR|λ|,m (x, .)k2 kLR|λ|,m (., y)k2 . By the analyticity assumption on e−tA we have supλ∈Σ(θ+ π2 ) kλ(λI + A)−1 kL(L2 ) < ∞. In particular, kLR|λ|,m (., y)k2 ≤ M kR|λ|,m (., y)k2 with some constant M independent of λ. We then obtain |Rλ,2m (x, y)| ≤ kR|λ|,m (x, .)k2 kR|λ|,m (., y)k2
(9)
We now estimate kR|λ|,m (x, .)k2 . Using heat kernel bound (4), we have R | R|λ|,m (x, z)|2 dµ(z) X Z √ C −c |λ|d(x,z) dµ(z) ≤ e |λ|2m v(x, √1 )2 X |λ|
∞ Z X C = |λ|2m v(x, √1 )2 k=0 { √k |λ|
≤
√
e
|λ|
C |λ|2m v(x, √1 )2 |λ|
−c
|λ|d(x,z)
dµ(z)
k+1 ≤d(x,z)≤ √ } |λ|
∞ X
k+1 v(x, p )e−ck |λ| k=0
∞
X C (k + 1)n e−ck ≤ |λ|2m v(x, √1 ) k=0 |λ|
where we used (7) to obtain the last inequality. We now obtain from this and (9) that |Rλ,2m (x, y)| ≤
C |λ|2m v(x,
√1 )1/2 v(y, |λ|
√1 )1/2
.
(10)
|λ|
Now, Proposition 3.3 of [11] shows that the strong homogeneity property (7), together with the bound (10) for λ > 0, imply that √ C −c” |λ| |Rλ,2m (x, y)| ≤ e d(x, y) |λ|2m v(x, √1 )1/2 v(y, √1 )1/2 |λ|
for λ ∈ Σ(θ +
π ). 2
|λ|
This inequality and (7) imply (6) for λ ∈ Σ(θ + π2 ).
42
XUAN THINH DUONG AND EL MAATI OUHABAZ
We now show (γ) ⇒ (α). The proof is standard but we give the details for the sake of completeness. Using the inverse Laplace transform we have Z m−1 pt (x, y) = eλt Rλ,m (x, y)dλ 2πitm−1 ΓR where ΓR = {re−iα , r ≥ R} ∪ {Reiα , |φ| ≤ α} ∪ {reiα , r ≥ R} := Γ1 ∪ Γ2 ∪ Γ3 and where α ∈ ( π2 , ν + π2 ) is a given constant and R = 2 max( 1t , d (x,y) ). t2 Using the assertion 3, we can write that for some constants C, c, c0 independent of R Z √ C e<λt p |pt (x, y)| ≤ m−1 e−c |λ|d(x,y) d|λ| (11) m t |λ|−1 ) ΓR |λ| v(x, The doubling property (7) implies that for |λ| ≥ R ≥ t−1 , p √ n v(x, t) ≤ M (|λ|t) 2 v(x, |λ|−1 ). Hence, 1 tm−1
R
e<λt √ Γ1 ∪Γ3 |λ|m v(x, |λ|−1 )
e−c
√
|λ|d(x,y)
(12)
d|λ|
Z ∞ √ n C 0 √ ≤ (λt)−m+ 2 e−c λt e−c λd(x,y) dλ v(x, t) R Z ∞ √ 0 n c0 C −c Rd(x,y) − c2 tR √ e e ≤ s−m+ 2 e− 2 s ds v(x, t) 1 √ c0 C √ e−c Rd(x,y) e− 2 tR ... ≤ v(x, t) 2 (x,y)
Using the fact that R = max( 1t , d
t2
), the last term is dominated by
d2 (x,y) C √ e−c t . v(x, t)
R Again by (12) we can bound the third term ( i.e., Γ2 ) in the right hand side of (11) by Z √ n C 0 √ (Rt)−m+ 2 ec Rt e−c Rt td|λ|. v(x, t) |λ|=R This term is clearly dominated by √ n C √ (Rt)−m+ 2 +1 e−c Rt v(x, t) which gives the desired bound on pt (x, y). ♦
GAUSSIAN UPPER BOUNDS FOR HEAT KERNELS
43
Proof of Theorem 2. Suppose that b ∈ L∞ (X, µ, C) satisfies (1) and that −bA generates a bounded analytic semigroup on L2 (X, µ). For any λ > 0 we write (λI + bA)−1 = (λb−1 + A)−1 b−1
(13)
Our aim is to prove the following pointwise inequalities which is valid a.e. for all f ∈ L2 (X, µ) |(λI + bA)−1 f | ≤ kb−1 k∞ (λc0 I + A)−1 |f |
(14)
where c0 = kbkδ 2 . The proof of (14) was given in [10] but we repeat it ∞ here to keep this paper self sufficient. We first observe that the positivity of pt (x, y) implies that the resolvent (λc0 I + A)−1 is a positivity preserving operator for λ > 0. This is also the case for the operators (sI + λ<( 1b ) + A)−1 for all s, λ > 0 as a consequence of the Trotter product formula (see [17]) t
1
1
t
e−t(λ<( b )+A) f = limn→∞ (e− n λ<( b ) e− n A )n f, ∀f ∈ L2 (X) ( in [17], the formula is given for contraction semigroups but it applies 1 in this situation since both semigroups e−t<( b ) and e−tA are contractions for the equivalent norm kf k∗ := supt≥0 ke−tA |f |k2 ). Note that by this formula we have the pointwise estimate 1
e−t(λ<( b )+A) |f | ≤ e−tA |f |, ∀t > 0, and f ∈ L2 (X) from which it follows that Z ∞ 1 1 −1 (sI + λ<( ) + A) = e−st e−t(λ<( b )+A) dt b 0 exists for all s, λ > 0. λ Using again the Trotter product formula for e−t( b +A) , we deduce that λ
1
|e−t( b +A) f | ≤ e−t(λ<( b )+A) |f | which implies 1 λ + A)−1 b−1 f | ≤ kb−1 k∞ (sI + λ<( ) + A)−1 |f |. b b 1 δ Since <( b ) ≥ kbk2 := c0 , it follows from the Trotter product formula ∞ that λ |(sI + + A)−1 b−1 f | ≤ kb−1 k∞ (sI + λc0 I + A)−1 |f |. b λ −1 Since ( b +A) and (λc0 I +A)−1 exist we conclude from this inequality that λ |( + A)−1 b−1 f | ≤ kb−1 k∞ (λc0 I + A)−1 |f | b |(sI +
44
XUAN THINH DUONG AND EL MAATI OUHABAZ
which is the desired inequality [14]. We now choose an integer m > n where n is the constant in the homogeneity property (7). Iterate (13) m−times, we obtain −m |(λI + bA)−m f | ≤ kb−1 km |f | . ∞ (λc0 I + A)
(15)
By assumption, pt (x, y) satisfies (4) hence by Theorem 1, (λc0 I√+ A)−m has a kernel Rλ,m (x, y) which satisfies (6). In particular, v(., λ−1 ) × (λc0√ I + A)−m is bounded from L1 into L∞ . Estimate (15) implies that v(., λ−1 )(λI + bA)−m is bounded from L1 into L∞ , so it is given by a kernel. This implies that (λI + bA)−m is given by a kernel. Again by (15), this kernel satisfies (6). By Theorem 1 we conclude that the kernel kt (x, y) of e−tbA satisfies (5). ♦ References [1] W. Arendt, Gaussian estimates and interpolation of the spectrum in Lp , Differential and Integral Equations 7 (1994) 1153-1168 [2] W. Arendt and A.F.M. ter Elst, Gaussian estimates for secondorder elliptic operators with boundary conditions, J. Op. Theory ( to appear) [3] P. Auscher, Regularity theorems and heat kernels for elliptic operators, J. Math. London Soc., 54 (1996) 284-296 [4] P. Auscher, A. McIntosh, Ph. Tchamitchian, Heat kernels of second-order complex elliptic operators and applications, J. F. Anal 152 (1998) 22-73 [5] P. Bauman, Equivalence of the Green’s functions for diffusion operators in Rn : a counter example, Proc. Amer. Math. Soc 91 (1984) 64-68 [6] P. Bauman, Positive solutions of elliptic equations in nondivergence form and their adjoints, Ark. Mat. 22 (1984) 153-173 [7] Th. Coulhon and X.T. Duong, Maximal regularity and kernel bounds: observations on a theorem by Hieber and Pr¨ uss, Diff. Integ. Eq. (to appear) [8] E.B. Davies, Heat Kernels and Spectral Theory, Cambridge Univ. Press 1989 [9] E.B. Davies, Lp spectral independence and L1 analyticity, J. London Math. Soc. (2)52 (1995) 177-184 [10] X.T. Duong and E.M. Ouhabaz, Complex multiplicative perturbations of elliptic operators: Heat kernel bounds and holomorphic functional calculus, Diff. Integ. Eq ( to appear) [11] X.T. Duong and D.W. Robinson, Semigroup kernels, Poisson bounds and holomorphic functional calculus, J. Funct. Analysis 142 (1996) 89128. ¨ss, Heat kernels and maximal Lp -Lq estimates for [12] M. Hieber, J. Pru parabolic evolution equations, Preprint 1996 [13] T. Kato, Remarks on Schr¨odinger operators with vector potentials, Int. Equ. Op. Th. 1 (1979) 103-113
GAUSSIAN UPPER BOUNDS FOR HEAT KERNELS
45
[14] A. McIntosh and A. Nahmod, Heat kernel estimates and functional calculi of −b∆, Preprint 1996 [15] E.M. Ouhabaz, Gaussian estimates and holomorphy of semigroups, Proc. Amer. Math. Soc. Vol. 123, N.5 (1995) 1465-1474 [16] E.M. Ouhabaz, Heat kernels of multiplicative perturbations: H¨older estimates and Gaussian lower bounds, Preprint 1998 [17] M. Reed and B. Simon, Methods of modern mathematical physics, Vol. 2, Academic Press, 1978. [18] N.Th. Varopoulos, L. Saloff-Coste and Th. Coulhon, Analysis and Geometry on Groups, Cambridge Univ. Press 1992 School of Mathematics, Physics, Computing and Electronics, Macquarie University N.S.W. 2109 Australia E-mail address: [email protected] ´matiques Applique ´es, Universite ´ de Equipe d’Analyse et de Mathe ´ ´ Marne-La-Vallee. Cite Descartes, 77454 Champs-sur-Marne, Marne´e cedex 2, France. La-Valle E-mail address: [email protected]
WEAK TYPE (1,1) ESTIMATES OF MAXIMAL TRUNCATED SINGULAR OPERATORS XUAN THINH DUONG AND LIXIN YAN Abstract. Let X be a space of homogeneous type and T a singular integral operator which is bounded on L2 (X ). We give a sufficient condition on the kernel of T so that the maximal truncated operator T∗ , which is defined by T∗ f (x) = sup>0 |T f (x)|, to be of weak type (1, 1). Our condition is weaker than the usual H¨ ormander type condition. Applications include the dominated convergence theorem of holomorphic functional calculi of linear elliptic operators on irregular domains.
1. Introduction and main theorem Let us consider a space of homogeneous type (X , d, µ) which is a set X endowed with a distance d and a non-negative Borel measure µ on X such that the doubling condition µ(B(x, 2r)) ≤ cµ(B(x, r)) < ∞ holds for all x ∈ X and r > 0, where B(x, r) = {y ∈ X : d(x, y) < r}. A more general defintion can be found in [CW, Chapter 3]. The doubling property implies the following strong homogeneity property, (1.1)
µ(B(x; λr)) ≤ cλn µ(B(x; r))
for some c, n > 0 uniformly for all λ ≥ 1. The parameter n is a measure of the dimension of the space. There also exist c and N, 0 ≤ N ≤ n so that N d(x, y) (1.2) µ(B(y; r)) ≤ c 1 + µ(B(x; r)) r uniformly for all x, y ∈ X and r > 0. Indeed, the property (1.2) with N = n is a direct consequence of triangle inequality of the metric d and the strong homogeneity property. In the case of Euclidean spaces Rn and Lie groups of polynomial growth, N can be chosen to be 0. 1
2000 Mathematics Subject Classification: 42B20, 47F, 47G10 Both authors are partially supported by a grant from Australia Research Council, and the second author is also partially supported by NSF of China. 2
46
WEAK TYPE (1,1) ESTIMATES
47
We consider the following “generalised approximations to the identity” which was introduced in [DM]. DEFINITION 1.1. A family of operators {At , t > 0} is said to be a “gen-
eralised approximation to the identity” if, for every t > 0, At is represented by kernels at (x, y) in the following sense: for every function f ∈ Lp (X ), p ≥ 1, Z At f (x) = at (x, y)f (y)dµ(y); X
and the following condition holds: (1.3)
|at (x, y)| ≤ ht (x, y) =
1 s(d(x, y)m t−1 ), µ(B(x; t1/m ))
where m is a positive fixed constant and s is a positive, bounded, decreasing function satisfying lim rn+N + s(rm ) = 0
(1.4)
r→∞
for some > 0, where n and N are two constants in (1.1) and (1.2). The operators we are going to consider henceforth were introduced in [DM]. They are defined in the following way. (1.5) T is a bounded operator on L2 (X ) with an associated kernel k(x, y) such that for f ∈ L∞ 0 (X ), Z T (f )(x) = k(x, y)f (y)dµ(y), for µ −almost every x 6∈ suppf. X
(1.6) There exists a “generalised approximation of the identity” {At , t > 0} such that T At have associated kernels kt (x, y) and there exist constants c1 , c2 > 0 so that Z |k(x, y) − kt (x, y)|dµ(x) ≤ c2 , for all y ∈ X . d(x,y)≥c1 t1/m
(1.7) There exists a “generalised approximation of the identity” {Bt , t > 0} such that Bt T have kernels Kt (x, y) which satisfy |Kt (x, y)| ≤ c4
1 , µ(B(x; t1/m ))
when d(x, y) ≤ c3 t1/m
and |Kt (x, y)−k(x, y)| ≤ c4
1 tα/m , when d(x, y) ≥ c3 t1/m , µ(B(x; d(x, y))) d(x, y)α
for some constants c3 , c4 , α > 0.
48
XUAN THINH DUONG AND LIXIN YAN
We assume that T is an operator satisfying (1.5), (1.6) and (1.7). The maximal operator T∗ is the supremum of the truncated integrals, namely, Z T∗ f (x) = sup |T f (x)| = sup k(x, y)f (y)dµ(y) . >0 >0 d(x,y)≥
It is proved in [DM] that if T verifies (1.5) and (1.6), then it is of weak type (1,1) and of strong type (p, p) for 1 < p ≤ 2. In addition, if (1.7) is also satisfied, the operator T is bounded on Lp (X ) for all 1 < p < ∞. Furthermore, Theorem 3 [DM] shows that T∗ is bounded on Lp (X ), 1 < p < ∞. Implicitly in the proof we can find the following Cotlar type inequality: T∗ f (x) ≤ CM (T f )(x) + CM f (x), where M is the Hardy-Littlewood maximal function. Hence, boundedness of T∗ follows from boundedness of T and M . The following is the main result of this paper. THEOREM 1.2. Let T be an operator satisfying the assumptions (1.5)
and (1.7). Also assume the following condition (1.8): there exists a “generalised approximation of the identity” {At , t > 0} so that the kernels (K,t (x, y) − K (x, y)) of the operators (B T At − B T ) satisfy Z (1.8) sup |K,t (x, y) − K (x, y)|dµ(x) ≤ C
d(x,y)≥βt1/m
for some constants C and β, and for all y ∈ X . Then, (i) the maximal truncated operator T∗ is bounded on Lp (X ) for 1 < p < ∞. (ii) When p = 1, T∗ is of weak-type (1, 1), that is, µ({x : |T∗ f (x) > α}) ≤
C kf k1 , for all α > 0. α
NOTE: (i) Comparing the above Theorem 1.2 with Theorem 3 of [DM],
the assumption (1.8) is stronger than (1.6), but we obtain new endpoint estimates, i.e. the weak type (1,1) estimates in (ii). (ii) Theorem 1.2 improves the classical results of Calder´on-Zygmund operators. See [St, Chapter 1, Corollary 2] for Euclidean spaces X = Rn , and [CW] for spaces of homogeneous type. Let us note that there is no regularity assumptions in the space variables. In comparison with the classical Calder´on-Zygmund operators, the H¨ormander type inequalities are replaced by (1.7) and (1.8) which involved the “generalised approximations of the identity”. In fact, for suitable “generalised
WEAK TYPE (1,1) ESTIMATES
49
approximations of the identity”, it is proved in [DM] that conditions (1.6) and (1.7) are weaker than the usual assumptions of Calder´onZygmund operators (Proposition 2, [DM]). We also show that our condition (1.8) is actually a consequence of the condition (1.6) (Proposition 2.1). As applications, we get the dominated convergence theorem of holomorphic functional calculi of linear elliptic operators on irregular domains. 2. Proof of Theorem 1.2 We first prove (ii). For a fixed > 0, one writes T u(x) = Bm T u(x)− (Bm T − T )(u)(x). Since the class of operators Bt satisfies the conditions (1.3) and (1.4), we have |Bm T u(x)| ≤ cM (|T u(x)|)
(2.1)
where c is a constant independent of . Similarly to the proof of Theorem 3 in [DM], using the condition (1.7) we also have sup |(Bm T − T )u(x)| ≤ cM (|u|(x)).
(2.2)
>0
Theorem 1.2 then follows from (2.2) if we can prove that the operator T∗B u(x) = sup |Bm T u(x)|
is of weak-type (1, 1). Following the idea of Theorem 1 of [DM], we first use the Calder´on-Zygmund decomposition to decompose an integrable function into “good” and “bad” parts (see, for example, [CW]). Given f ∈ L1 (X ) ∩ L2 (X ) and α > kf k1 (µ(X ))−1 , then there exist a constant c independent of f and α, and a decomposition X f =g+b=g+ bi , i
so that (a) |g(x)| ≤ cα for all almost x ∈ X ; (b) there exists a sequence of balls Qi so that the support of each bi is contained in Qi and Z |bi (x)|dµ(x) ≤ cαµ(Qi ); X R P (c) µ(Qi ) ≤ cα−1 X |f |dµ(x); i
(d) each point of X is contained in at most a finite number N of the balls Qi . Note that conditions (b) and (c) imply that kbk1 ≤ ckf k1 and hence that kgk1 ≤ (1 + c)kf k1 .
50
XUAN THINH DUONG AND LIXIN YAN
We have µ({x : |T∗B f (x)| > α}) ≤ µ({x : |T∗B g(x)| > α/2}) + µ({x : |T∗B b(x)| > α/2}). It follows from (2.1), (1.5) and boundedness of the Hardy-Littlewood maximal function that T∗B is bounded on L2 (X ). Since |g(x)| ≤ cα, we obtain (2.3) µ({x : |T∗B g(x)| > α/2}) ≤ 4α−2 kT∗B gk22 ≤ cα−2 kgk22 ≤ cα−1 kf k1 . Concerning the “bad” part b(x), we temporarily fix a bi whose support is contained in Qi , then choose ti = rim where m is the constant appearing in (1.3), and ri is the radius of the ball Qi . We then decompose X sup B T bi (x)
i
X X ≤ sup B T Ati bi (x) + B T (I − Ati )bi (x) .
i
i
It follows from the decay assumption (1.3) that
X
X 1/2
1/2 Ati bi ≤ cα µ(Qi ) ≤ cα1/2 kf k1 .
2
i
i
See details in the proof of estimate (10) in [DM]. Combining this with L2 -boundedness of T∗B , we have
X
2 P
(2.4) µ({x : sup |B T i Ati bi (x)| > α/4}) ≤ 16α−2 T∗B Ati bi
2
i
X
2
≤ cα−2 Ati bi i
≤
2
c kf k1 . α
On the other hand µ({x : sup |B T
P
− Ati )bi (x)| > α/4}) X X4Z sup |B T (I − Ati )bi (x)|dµ(x), ≤ µ(Bi ) + α (Bi )c i i
i (I
where (Bi )c denotes the complement of Bi which is the ball with the same centre yi as that of the ball Qi in the Calder´on-Zygmund decomposition but with radius increased by a factor of (1+c1 ), where c1 is the
WEAK TYPE (1,1) ESTIMATES
51
constant in (1.6). Because of property (c) of the Calder´on-Zygmund decomposition and doubling volume property of X , we have X X (2.5) µ(Bi ) ≤ c µ(Qi ) ≤ cα−1 kf k1 . i
i
Using assumption (1.8), we have Z sup |B T (I − Ati )bi (x)|dµ(x) (Bi )c Z Z ≤ sup (K (x, y) − K,ti (x, y))bi (y)dµ(y) dµ(x) (Bi )c X Z Z ≤ kbi (y)k sup sup |K (x, y) − K,ti (x, y)|dµ(x) dµ(y) X
≤ Ckbi k1 ,
y
d(x,y)≥cti 1/m 1/m
because B(y; cti
) ⊂ Bi .
Therefore (2.6) Z X1 X C sup |B T (I − Ati )bi (x)|dµ(x) ≤ Cα−1 kbi k1 ≤ kf k1 . α (Bi )c α i i Combining the above estimates (2.2), (2.3), (2.4), (2.5) and (2.6), we have for any α > kf k1 (µ(X ))−1 , C µ({x : |T∗ f (x)| > α}) ≤ kf k1 . α If X is unbounded, the proof is done because the former inequality holds for every α > 0. Otherwise, we have to consider what happens for 0 < α ≤ kf k1 (µ(X ))−1 . Since X is bounded we can write X = B(x0 , r) for some r > 0. We conclude C µ({x : |T∗ f (x)| > α}) ≤ µ(X ) ≤ kf k1 α for any α > 0. We now prove (i). For any 1 < p ≤ 2, Lp -boundedness of T∗ follows from the Marcinkiewicz interpolation theorem. Using a standard duality argument, T∗ is proved to be a bounded operator on Lp (X ) for all 2 < p < ∞. The proof of Theorem 1.2 is complete. In the next proposition, we show that, for suitable chosen Bt , our condition (1.8) is actually a consequence of condition (1.6). PROPOSITION 2.1. Let T be a bounded linear operator on L2 (X ) with
kernel k(x, y). Assume there exists a “generalised approximation of
52
XUAN THINH DUONG AND LIXIN YAN
the identity” {At , t > 0} so that the kernels kt (x, y) of T At satisfy the condition (1.6), i.e. there exist constant c and δ > 1 so that Z (2.7) |k(x, y) − kt (x, y)|dµ(x) ≤ c d(x,y)≥δt1/m
for all y ∈ X . Then, there exists a “generalised approximation of the identity” {Bt , t > 0} which is represented by kernels bt (x, y) in the following sense: for any f ∈ Lp (X ), p ≥ 1, Z Bt f (x) = bt (x, y)f (y)dµ(y), X
so that the kernels (K,t (x, y)−K (x, y)) of the operators (B T At −B T ) satisfy Z |K,t (x, y) − K (x, y)|dµ(x) ≤ C
sup
d(x,y)≥βt1/m
for some constants C and β, and for all y ∈ X . Proof. Choose δ > 1 and let β = 3δ/2. For any > 0, we choose b (x, z) = 0 when d(x, z) ≥ (δ/2)t1/m . Then, for x, y ∈ X so that Z K,t (x, y) = b (x, z)kt (z, y)dµ(z), X Z b (x, z)k(z, y)dµ(z). and K (x, y) = X
For all y ∈ X , Z |K,t (x, y) − K (x, y)|dµ(x) d(x,y)≥βt1/m Z Z ≤ |b (x, z)||kt (z, y) − k(z, y)|dµ(x)dµ(z) d(x,y)≥βt1/m X Z Z |b (x, z)|dµ(x) × |kt (z, y) − k(z, y)|dµ(z) ≤ sup z∈X X d(z,y)≥δt1/m Z ≤ c1 |kt (z, y) − k(z, y)|dµ(z) d(z,y)≥δt1/m
≤ C, where the last inequality follows from (2.7) and the third inequality is using the estimate Z Z |b (x, z)|dµ(x) ≤ h (x, z)dµ(z) ≤ c1 . X
X
WEAK TYPE (1,1) ESTIMATES
53
As a consequence of the boundedness of the maximal truncated operator T∗ , we obtain pointwise almost everywhere convergence of lim→0 T f (x). More precisely, we have the following corollary. COROLLARY 2.2. Assume that the operator T satisfies the conditions of
Theorem 1.2. Assume that the kernel k(x, y) of T satisfies the estimate |k(x, y)| ≤ c(µ(B(x; d(x, y)))−1 . Then there exist a sequence of positive functions j (x) such that lim j (x) = 0, and a function m ∈ L∞ (X ) such that for f ∈ Lp (X ), 1 ≤ j→∞ p < ∞, Z T f (x) = m(x)f (x) + lim k(x, y)f (y)dµ(y) j→∞
|x−y|≥j (x)
for almost every x ∈ X . Proof. Corollary 2.2 follows from a standard argument of proving the existence of almost everywhere pointwise limits as a consequence of the corresponding maximal inequality. See, for example, [CM, Chapter 7, Theorem 6] for Euclidean spaces X = Rn , and [CW] for spaces of homogeneous type. REMARK 2.3.
As in Section 3 of [DM], Theorem 1.2 and Corollary 2.2 can be modified so that they are still true when the space of homogeneous type X is replaced by one of its measurable subsets Ω. In this sense, it is sufficient that condition (1.3) on the upper bound ht (x, y) of the kernel at (x, y) is replaced by ht (x, y) = (µ(B X (x; t1/m )))−1 s(d(x, y)m t−1 ), where B X (x; t1/m ) is the ball of centre x, radius t1/m in the space X . For the details, see Section 3, [DM]. 3. Applications: Holomorphic functional calculi of linear elliptic operators We first review some definitions regarding the holomorphic functional calculus as introduced by McIntosh [Mc]. Let 0 ≤ ω < π be given. Then Sω = {z ∈ C : |argz| ≤ ω} ∪ {0} denotes the closed sector of angle ω and Sω0 denotes its interior, while S˙ ω = Sω \{0}. An operator L on some Banach space E is said to be
54
XUAN THINH DUONG AND LIXIN YAN
of type ω if L is closed and densely defined, σ(L) ⊂ Sω , and for each θ ∈ (ω, π] there exists a constant Cθ such that |η| k(ηI − L)−1 kL(E) ≤ Cθ , η ∈ −S˙ π−θ . If µ ∈ (0, π], then H∞ (Sµ0 ) = {f : Sµ0 → C; f is holomorphic and ||f ||∞ < ∞}, where ||f ||H∞ = sup{|f (z)| : z ∈ Sµ0 }. In addition, we define |z|s 0 0 Ψ(Sµ ) = g ∈ H∞ (Sµ ) : ∃s > 0, ∃c ≥ 0 : |g(z)| ≤ c . 1 + |z|2s If L is of type ω and g ∈ Ψ(Sµ0 ), we define g(L) ∈ L(E) by Z 1 (ηI − L)−1 g(η)dη, (3.1) g(L) = − 2πi Γ where Γ is the contour {ξ = re±iθ : r ≥ 0} parametrised clockwise around Sω , and ω < θ < µ. If, in addition, L is one-one and has dense range and if g ∈ H∞ (Sµ0 ), then (3.2)
f (L) = [h(L)]−1 (f h)(L),
where h(z) = z(1 + z)−2 . It can be shown that g(L) is a well-defined linear operator in E and that this definition is consistent with definition (3.2) for g ∈ Ψ(Sµ0 ). The definition of g(L) can even extended to encompass unbounded holomorphic functions; see [Mc] for details. L is said to have a bounded holomorphic functional calculus on the sector Sµ if ||g(L)||L(E) ≤ N ||g||∞ for some N > 0, and for all g ∈ H∞ (Sµ0 ). Assume that Ω is a measurable subset of a space of homogeneous type (X , d, µ). Let L be a linear operator on L2 (Ω) with ω < π/2 so that (−L) generates a holomorphic semigroup e−zL , 0 ≤ |Arg(z)| < π/2 − ω which possesses the following two properties: (3.3) The holomorphic semigroup e−zL , |Arg(z)| < π/2 − ω is represented by kernels az (x, y) which satisfy, for all θ > ω, an upper bound |az (x, y)| ≤ cθ h|z| (x, y) for x, y ∈ Ω, and |Arg(z)| < π/2 − θ, where h|z| is defined on X × X by (1.3). (3.4) The operator L has a bounded holomorphic functional calculus in L2 (Ω). That is, for any ν > ω and g ∈ H∞ (Sν0 ), the operator g(L) satisfies ||g(L)f ||2 ≤ cν ||g||∞ kf k2 . Applying Theorem 1.2, Corollary 2.2 and Remark 2.3, we have
WEAK TYPE (1,1) ESTIMATES
55
THEOREM 3.1. Let L be an operator verifying the assumptions (3.3) and (3.4). Assume that for g ∈ H∞ (Sν0 ), the kernel k(x, y) of g(L) satisfies the estimate
(3.5)
|k(x, y)| ≤ c(µ(B(x; d(x, y)))−1 ,
for all x, y ∈ Ω.
If we denote T = g(L), then (i) If 1 < p < ∞, then kT∗ f kp ≤ Ckgk∞ kf kp . (ii) If f ∈ L1 (X ), then the map f → T∗ f is weak type (1, 1). (iii) There exist a sequence of positive functions j (x) such that lim j (x) = 0, and a function m(x) ∈ L∞ (Ω) such that for f (x) ∈
j→∞
Lp (Ω), 1 ≤ p < ∞, Z T f (x) = m(x)f (x) + lim
j→∞
K(x, y)f (y)dµ(y) |x−y|≥j (x)
for almost every x ∈ Ω. Proof. We follow an idea of Theorem 6 in [DM]. Choose operators Bt = At = e−tL . As in Theorem 6 of [DM], there exist some constants c, c1 , α > 0 such that the kernels Kt (x, y) of Bt T satisfy 1 |Kt (x, y)| ≤ c , X µ(B (x; t1/m )) for all x, y ∈ Ω such that when d(x, y) ≤ c1 t1/m ; tα/m 1 |Kt (x, y) − k(x, y)| ≤ c µ(B X (x; d(x, y))) d(x, y)α for all x, y ∈ Ω such that when d(x, y) ≥ c1 t1/m . Using the methods of Theorems 5 and 6 [DM], it is not difficult to check that the kernels (K,t (x, y) − K (x, y)) of the operators (B T At − B T ) satisfy Z (3.6) sup |K,t (x, y) − K (x, y)|dµ(x) ≤ C
d(x,y)≥βt1/m
for some constants C and β, and for all y ∈ Ω. For the proof of (3.6), we leave it to readers. Then, Theorem 3.1 follows from Remark 2.3. REMARK 3.2.
The condition (3.5) is satisfied by large classes of linear operators on Rn or a domain Ω of Rn without any condition on smoothness of the boundary of Ω. For example, if the function ht (x, y) of (1.3) is bounded above by the Gaussian bounds
56
XUAN THINH DUONG AND LIXIN YAN
ct−n/2 exp{−α|x − y|2 /t} for some α > 0, or by the Poisson bounds ct , (t2 + |x − y|2 )(n+1)/2 then (3.5) is a direct result from straightforward integration. One example of an operator L which possesses Gaussian bounds on its heat kernel is the Schr¨odinger operator with potential V , defined by L = −4 + V (x), where V is a nonnegative function on Rn . See, Lecture 7 in [ADM]. Another example of an operator L on such a domain, which possesses Gaussian bounds on its heat kernel, is the Laplacian on an open subset of Rn subject to Dirichlet boundary conditions. More general operators on open domain of Rn which possess Gaussian bounds can be found in [AE], [DM]. References [ADM] D. Albrecht, X.T. Duong and A. McIntosh, ‘Operator theory and harmonic analysis’, Instructional Workshop on Analysis and Geometry (Proc. Centre Math. Analysis, 34, A.N.U., Canberra, 1996) pp 77-136. [AE] W. Arendt and A.F.M. ter Elst, Gaussian estimates for second order elliptic operators with boundary conditions, Reports on Applied and Numerical Analysis, Eindhoven University of Technology, 1995. [CM] R. Coifman and Y. Meyer, Ondelettes et Operateurs, II, Hermann, 1990, 19991. [CW] R.R. Coifman, G. Weiss, Analyse harmonique non-commutative sur certains espaces homog`es, Lecture Notes in Mathematics, 242 Springer-Verlag, Berlin, 1971. [DM] X.T. Duong and A. McIntosh, Singular integral operators with non-smooth kernels on irregular domains, Rev. Mat. Iberoamericana, 15 (1999), 233265. [Mc] A. McIntosh, Operators which have an H∞ -calculus, Miniconference on Operator Theory and Partial Differential Equations, 1986, Proceedings of the Centre for Mathematical Analysis, ANU, Canberra, 1986, 210-231. [St] E.M. Stein, Harmonic Analysis: Real variable methods, orthogonality and oscillatory integrals, Princeton Univ. Press, 1993. DEPARTMENT OF MATHEMATICS, MACQUARIE UNIVERSITY, NSW 2109, AUSTRALIA E-mail address: [email protected] DEPARTMENT OF MATHEMATICS, MACQUARIE UNIVERSITY, NSW 2109, AUSTRALIA; DEPARTMENT OF MATHEMATICS, ZHONGSHAN UNIVERSITY, GUANGZHOU, 510275, P.R. CHINA E-mail address: [email protected]
L2 BOUNDS FOR NORMAL DERIVATIVES OF DIRICHLET EIGENFUNCTIONS ANDREW HASSELL AND TERENCE TAO
Abstract. Suppose that M is a compact Riemannian manifold with boundary and u is an L2 -normalized Dirichlet eigenfunction with eigenvalue λ. Let ψ be its normal derivative at the boundary. Scaling considerations lead one to expect that the L2 norm of ψ will grow as λ1/2 as λ → ∞. We sketch proofs of an upper bound of the form kψk22 ≤ Cλ for any Riemannian manifold, and a lower bound cλ ≤ kψk22 provided that M has no trapped geodesics (see the main Theorem for a precise statement). Here c and C are positive constants that depend on M , but not on λ. Full details will appear in [3].
1. Introduction Let M be a smooth compact Riemannian manifold with smooth boundary ∂M = Y (for example, the closure of a smooth bounded domain in Euclidean space). Let H = −∆M be minus the Dirichlet Laplacian on M . As is well known, H has discrete spectrum 0 < λ1 < λ2 ≤ λ3 · · · → ∞. Let uj be an L2 -normalized eigenfunction corresponding to λj , and let ψj be the normal derivative of uj at the boundary. In this paper we consider the following question: do there exist constants c and C, depending on M but not on j, such that cλj ≤ kψj k2L2 (Y ) ≤ Cλj ? This question was posed by Ozawa in [6]; he showed that a weaker version of this statement, obtained by summing over all eigenvalues in [0, λ], is true. A. H. is supported by an Australian Research Council Fellowship. T. T. is a Clay Prize Fellow and is supported by the Packard Foundation. 57
58
ANDREW HASSELL AND TERENCE TAO
In this paper, we sketch proofs that the upper bound is always true on Riemannian manifolds, while the lower bound holds provided a condition of ‘no trapped geodesics’ holds. In addition, we give several examples to illustrate the link between failure of the geodesic condition and failure of the lower bound. Full details of proofs are given in [3]. We acknowledge helpful conversations with Alan McIntosh, Steve Zelditch and Johannes Sj¨ostrand. 2. Examples We begin by considering several examples. These examples show in particular that the lower bound does not always hold. Example 1 — the disc. Let M = {x ∈ R2 | |x| < a} for some a > 0. In this case we have an equality Z 2λj ψj2 (θ) dθ = . a S1 This follows easily from an identity for eigenfunctions due to Rellich, which we discuss below. Example 2 — the rectangle. Let M = [0, a]×[0, b], where a ≤ b. Then it is a simple matter to write down the eigenfunctions, by separating variables. A computation gives 4 4 λ ≤ kψk22 ≤ λ, b a and these bounds are the best possible. 1 Example 3 — the cylinder. Let M = [0, π] × S2π , the product of an interval of length π with a circle of length 2π. Then eigenfunctions take the form sin(mx)einθ . The upper bound
4 λ π holds, but by holding m fixed and sending n to infinity, we see that the best lower bound is O(1). kψk22 ≤
Example 4 — the hemisphere. Let M be the hemisphere M = {x ∈ R3 | |x| = 1, x · (0, 0, 1) ≥ 0}. In this case, the eigenfunctions are given by those spherical harmonics which are odd under reflection in the (x1 , x2 ) plane, namely, spherical
L2 BOUNDS FOR DIRICHLET EIGENFUNCTIONS
59
harmonics u = clm Ylm = clm eimφ Plm (cos θ),
λ = l(l + 1),
where −l ≤ m ≤ l and l −m is odd. (Here, we are using spherical polar coordinates where (x1 , x2 , x3 ) = (r sin θ cos φ, r sin θ sin φ, r cos θ).) Let us consider the case when m = l − 1. Then the eigenfunction in this case is ul = cei(l−1)θ cos θ(sin θ)l−1 . A short computation gives kψk22 ∼ l3/2 ∼ λ3/4 ,
l → ∞,
for this class of eigenfunctions. Hence there is no nontrivial lower bound for the hemisphere. 3. Main theorem The upper bound holds in all the examples in the previous section, but the lower bound fails for the last two. To see more heuristically why the lower bound fails, it helps to consider a more dynamic picture, by considering the wave equation ∂ 2 v(x, t) = −Hv ∂t2 on the cylinder (Example 3). If u is an eigenfunction with eigenvalue √ i λt λ, then v = e u is a solution to the wave equation. Thus in the case of the cylinder, a particular solution to the wave equation is √
2i sin(mx)einθ ei
m2 +n2 t
√
= eimx einθ ei
m2 +n2 t
√
− e−imx einθ ei
m2 +n2 t
.
The wavefronts are at ±mx + nθ = constant, and energy moves ‘normal’ to the wavefronts. When we hold m fixed and send n to infinity, the energy is moving more and more along lines (that is, geodesics) where x is constant, and so does not ‘reach’ the boundary. Thus, the failure seems to be related to the existence of a family of ‘trapped’ geodesics in the interior of the manifold, which never reach the boundary. In the case of the hemisphere, there is no geodesic trapped in the interior, but any Riemannian extension N of M would have a trapped geodesic, namely the boundary of M . Note that the lower bound is not violated as severely for the hemisphere, which is consistent with it being a borderline case. Our main result is that the heuristic above is correct: Theorem 3.1. Let M be a smooth compact Riemannian manifold with boundary. Then the upper bound holds for some C independent of j.
60
ANDREW HASSELL AND TERENCE TAO
The lower bound holds provided that M can be embedded in the interior of a compact manifold with boundary, N , of the same dimension, such that every geodesic in M eventually meets the boundary of N . In particular, the lower bound holds if M is a subdomain of Euclidean space. In Section 7, we give some further examples which show that the degree of failure of the bounds is related to the extent to which geodesics are trapped. 4. Upper bound Our proof is based on the following Lemma which we call a Rellichtype estimate. Lemma 4.1. Let u be a Dirichlet eigenfunction of H. Then for any differential operator A, Z Z ∂u hu, [H, A]ui dg = (4.1) Au dσ. ∂ν M
Y
Proof. Let λ be the eigenvalue corresponding to u. We write [H, A] = [H − λ, A] and use the fact that (H − λ)u = 0 to write the integral over M as Z h(H − λ)u, Aui − hu, (H − λ)Aui dg. M
Then we use Green’s formula, and the fact that u vanishes at the boundary, to deduce (4.1). The upper bound is now easily deduced. Let us choose coordinates (r, y) near the boundary of M , where r is distance to the boundary (which is a smooth function for r < δ, for some sufficiently small δ > 0) and y are local coordinates on Y = ∂M , extended to be constant along geodesics perpendicular to the boundary. Let χ(r) be a smooth function which is supported in [0, δ/2] and with χ(0) = 1, and let A = χ(r)∂r . Then the right hand side of (4.1) becomes kψk22 , while the left hand side is bounded by Z C |∇u|2 + 1) = C(λ + 1), M
proving the upper bound. Remark. Notice that this argument actually gives a bound of C kHukL2 ({r≤}) = C λkukL2 ({r≤})
L2 BOUNDS FOR DIRICHLET EIGENFUNCTIONS
61
for any > 0. 5. Lower bound for Euclidean domains In the case of subdomains of Euclidean space, there is a very simple proof of the lower bound based on Lemma 4.1. We set n X ∂ A= xi . ∂x i i=1 Then [H, A] = 2H, so now the left hand side is 2λ, and we obtain the identity Z Z ∂u 2 ∂u Au dσ = ν · x dσ 2λ = ∂ν ∂ν Y
Y
which immediately implies the lower bound. (Rellich published this identity in 1940 [7].) 6. The lower bound in general It is considerably harder to prove the lower bound for general manifolds satisfying the ‘no trapped geodesic’ condition of Theorem 3.1. We use the method for Euclidean domains as a guide, and seek a first order operator A having a positive commutator with H. It is impossible to find a vector field A with this property, in general (we will show this later), so we look for a first order pseudodifferential operator. A is essentially determined by its principal symbol, a, a function on S ∗ M , the cosphere bundle of M . The principal symbol of i[H, A] is given by Vh (a), where Vh is the Hamilton vector field of the symbol of H. Recall that H is minus the Laplacian. It is well known that Vh is the generator of geodesic flow on S ∗ M . We want i[H, A] positive, which amounts to finding a smooth function a on S ∗ M which is increasing along all geodesics. Note that this is clearly impossible if trapped/periodic geodesics are present. Given the ‘no trapped geodesics’ condition on M in the theorem, such an A can easily be constructed. To show this, we first observe that, given a geodesic γ on M , one can construct an A on the larger manifold N ⊃ M , properly supported in the interior of N , such that the principal symbol of i[H, A] is everywhere nonnegative, and strictly positive in a neighbourhood of γ (at least that part of it lying over M ). To do this we simply define the symbol a of A to be linearly increasing along γ, extend it in a natural way, and then cut off. Finally, a compactness argument shows that one can add together finitely many
62
ANDREW HASSELL AND TERENCE TAO
such operators to produce one where the principal symbol of i[H, A] is strictly positive on S ∗ M . We note that it is easy to arrange that, in addition, A satisfies the transmission condition (see [5], section 18.2)); for the principal symbol a of A, this condition is simply that a is odd: a(x, −ξ) = −a(x, ξ). We then follow the strategy of the previous proof. First, we need a version of Lemma 4.1 which is valid for pseudodifferential operators. For first order pseudodifferential operators A satisfying the transmission condition, with symbol a, we have the following Lemma 6.1. Let u be a Dirichlet eigenfunction for H. Then Z Z Z ∂u ∂u 2 (6.1) hu, [H, A]ui dg = 2 Im Au dσ − c dσ, ∂ν ∂ν M
Y
Y
where c(y) = limρ→∞ ρ−1 a(0, y, ρ, 0). This is just as good as (4.1) for our purposes. We then follow the strategy of the previous proof. The left hand side causes few problems; it is rather easy to show that the left hand side is at least as big as a constant times λ. The right hand side, though, is more difficult. It is sufficient to show that √ (6.2) kAukL2 (Y ) ≤ C λ, since then we can estimate the first term on the right hand side of (6.1) by Z (6.3) hψ, Auidσ ≤ C()kψk22 + λ, Y
and the λ term may then be taken to the left hand side. The estimate (6.2) is nontrivial, since A is a nonlocal operator; the restriction of Au to the boundary depends on values of u in the interior of M . Our proof of (6.2) uses ideas from both harmonic analysis and microlocal analysis. There are two main ingredients. One is a uniform estimate for u near the boundary: Z (6.4) u2j dσ(y) ≤ Cλj r2 for all r ∈ [0, δ] Yr
where Yr is the set of points at distance r from the boundary, and both C and δ are independent of j. We prove this by looking at the quantity Z L(r) = u2 dσr . Yr
L2 BOUNDS FOR DIRICHLET EIGENFUNCTIONS
63
Here dσr is the measure on Yr induced by the Riemannian measure on M . The eigenfunction equation for u leads to the differential inequality (6.5)
L00 ≥
(L0 )2 − Cλ, L
for L(r), for some constant C depending only on the manifold M . This differential inequality implies exponential increase of L(r) if it ever happens that L0 (r)2 ≥ 2CλL(r). However, this would contradict the bound Z δ L(r)dr ≤ 1 0
which follows from the L2 normalization of u. Hence L0 (r)2 ≤ 2CλL(r), and from this we deduce (6.4). The second ingredient is expressing Au|Y as an integral of kernels Ar acting on the functions u|Yr . Here, As has kernel As (y, y 0 ) which is the restriction of the kernel of A to (r = 0, y, r0 = s, y 0 ). Then results of Boutet de Monvel [1] and Viˇsik and Eˇskin [8] give bounds on the L2 operator norm of Ar , acting from L2 (Yr ) to L2 (Y ). Their result is if Br is an operator of order −1 + k, satisfying the transmission condition, where k = 0, 1, 2, . . . , then there is a bound on the operator norm of Br of the form Cr−k , where C is independent of r. In particular, for k = 0, the Br are uniformly bounded as r → 0. Unfortunately, this result does not quite give the result directly, since if we combine the operator bound Cr−2 for Ar and (6.4) for u and integrate in r, we encounter a logarithmic divergence. However, a small modification of this approach does the trick. If we let H −1 denote the inverse of the operator H on N , with Dirichlet boundary conditions, then H −1 is a pseudodifferential operator of order −2 when localized away from the boundary of N . Moreover, it satisfies the transmission condition. Since Hu = ψδY + λu, we may write Au|Y = AH −1 ψδY + λu . The first term is given by (AH −1 )0 ψ which is a L2 -bounded operator applied to ψ. By the upper bound for ψ, we see that this term satisfies (6.2). For the second term, we write u as the sum of a ‘close’ part and a ‘far’ part with respect to the boundary, relative to the length scale λ−1/2 . We can integrate in r as described above to get the bound for the close part, while a similar argument works for the far part. This completes the proof of the lower bound.
64
ANDREW HASSELL AND TERENCE TAO
7. Periodic geodesics In this section, we explore the relationship between trapped geodesics and the failure of the lower bound by considering two examples. The first is a hyperbolic cylinder, that is, the manifold M = [−a, a] × Sθ1 , with metric g = dx2 + (cosh x)2 dθ2 . This has a single periodic geodesic, at x = 0, which is unstable (geodesics are always unstable in manfolds with negative curvature). Let > 0 be given, and let F be the set {|x| ≥ }; thus F excludes a neighbourhood of the trapped geodesic. Then Colin de Verdi`ere and Parisse showed that there is a sequence of normalized Dirichlet eigenfunctions ukj , kj → ∞ as j → ∞, such that Z 1 |ukj |2 dg ∼ , j → ∞. log λkj F Then, applying our upper bound argument, which only uses the norm of eigenfunctions in a neighbourhood of the boundary (see the remark at the end of Section 4), we see that for some C kψkj k22 ≤ C
λk j . log λkj
This shows that the lower bound is violated. However, we can actually show that this is the true order of growth of kψkj k22 . Indeed, we shall show that for every normalized eigenfunction u, (7.1)
kψk22 ≥ c
λ . log λ
To show (7.1), we first observe that for every eigenfunction u, we have Z C (7.2) |u|2 dg ≥ . log λ F This follows by looking at a basis of eigenfunctions ul,λ of the form ul,λ = eilθ (cosh x)−1/2 vl,λ . Then vl,λ satisfies 1 1 2 2 2 2 Dx + − (tanh x) + l (sech x) vl,λ = λvl,λ , 2 4 which we rewrite in the form (7.3) h2 Dx2 + V vl,λ = Evl,λ ,
L2 BOUNDS FOR DIRICHLET EIGENFUNCTIONS
65
with 1 h−2 = l2 + , 4
V (x) = (sech x)2 − 1,
E=
λ − l2 − 1/2 . l2 + 1/4
Then Theorem 20 of [2] shows that for |E| < C, (7.2) holds. For E > C, direct analysis of equation (7.3) shows that vl,λ has a uniform L∞ bound, independent of h and E. Thus, in this case (7.2) holds a fortiori. In the remaining case, E < −C < 0, the origin is in the classically forbidden region and the result follows immediately from Agmon-type exponential decay estimates. To complete the proof of (7.1), we construct an operator A which has a nonnegative commutator with H, and which is strictly positive in a neighbourhood of F . We are able to do this because of the special properties of geodesic flow on the hyperbolic cylinder. Letting G be the complement of F , and writing F = F+ ∪ F− for the two components of F , labelled according to the sign of x, a geodesic that passes from F+ , say, to G either stays over G for all subsequent time, or emerges into the region F− and eventually reaches the boundary of M ; it cannot happen that a geodesic starts in G, then moves into the set F , and returns to G. Thus, given a geodesic γ we can define an operator A with a symbol which is linearly increasing when γ is above the set F , and vanishing in some neighbourhood U ⊂⊂ G of the periodic geodesic. As before, we can (by compactness) find a finite number of such operators whose sum has the desired property. Thus, for a given eigenfunction u, let M (u) denote the quantity on the left hand side of (7.2). If we go back to (6.1), then we find that the left hand side is at least as big as cλM (u) − Cλ1/2 ≥ c0 λM (u). On the other hand, the argument above applied to Au shows that the right hand side is no bigger than Ckψk2 λ1/2 M (u)1/2 . We get the normalization factor M (u)1/2 (instead of 1) since all arguments are localized near the boundary of M . The combination of these two estimates yields (7.1). The second example we analyze is the spherical cylinder, that is, M = (−a, a) × S 1 , for 0 < a < 1, with metric g = dx2 + (cos x)2 dθ2 . This has a periodic geodesic at x = 0. However, in this case, the geodesic is stable, and indeed every nearby geodesic is periodic. Thus,
66
ANDREW HASSELL AND TERENCE TAO
this case is the opposite extreme where there is an open set of periodic geodesics. We shall see that, correspondingly, the lower bound is violated in an extreme fashion. This example may be analyzed in a similar way to the hyperbolic cylinder. We may separate variables, so there is a basis of eigenfunctions ul,λ of the form eilθ (cos x)−1/2 vl,λ . Then vl,λ satisfies the equation (7.4) h2 Dx2 + V vl,λ = Evl,λ , where now 1 λ − l2 + 1/2 h−2 = l2 − , V (x) = (sec x)2 − 1, E = . 4 l2 − 1/4 Here, V has a nondegenerate global mininum at x = 0. Let us consider a sequence of eigenfunctions ukj with lkj → ∞ and E(λkj , lkj ) ≤ E1 , where 0 < E1 < V (a), so that the boundary, |x| = a, is in the classically forbidden region. Then Agmon-type exponential decay estimates (see [4], chapter 3) hold, giving for some > 0 1/2
|ukj (x, θ)| ≤ Ce
−λk
j
,
|x| ≥ δ > 0,
for some > 0.
The upper bound argument then gives 1/2
kψkj k2 ≤ Ce
−0 λk
j
,
for some 0 > 0,
so we actually have exponential decrease, rather than O(λ1/2 ) increase, in the L2 norm of the normal derivative for any such sequence of eigenfunctions. 8. Vector fields are not enough Finally, we remark that the following example shows that one cannot expect to find a first order differential operator A having positive commutator with H. First we analyze what it means for a vector field to have a positive commutator with H. Let the symbol of A be ai (x)ξi . The Hamilton vector field of H is ξi ∂xi , so having a positive commutator requires that ∂aj (8.1) ξi ξj > 0 for |ξ| 6= 0. ∂xi Thus, the matrix ∂xi aj must be positive definite, or in other words a1 is increasing in direction e1 , etc. Now consider the manifold with boundary shown in the figure (the corners should be assumed to have been smoothed out so that it has
L2 BOUNDS FOR DIRICHLET EIGENFUNCTIONS
67
smooth boundary). In the figure, the topmost and bottom horizontal dotted lines, and the leftmost and rightmost dotted lines, are identified.
This manifold has no trapped geodesics. Suppose that there is a vector field A whose commutator with H has a positive symbol. Notice that the two points p and q are such that there are three geodesics from p to q: one in the direction e1 +e2 , one in the direction −e1 and one in the direction −e2 . Write A = a1 e1 + a2 e2 . Then a1 is increasing in direction e1 , and a2 is increasing in direction e2 . Hence a1 (p) > a1 (q), and a2 (p) > a2 (q). On the other hand, a1 + a2 is increasing in direction e1 +e2 , so this yields a1 (p)+a2 (p) < a1 (q)+a2 (q), which is a contradiction. References [1] L. Boutet de Monvel, Comportement d’un op´erateur pseudo-differentiel sur une vari´et´e ` a bord, J. Analyse Math. 17 (1966), 241-304. ´ [2] Y. Colin de Verdi`ere and B. Parisse, Equilibre instable en r´egime semiclassique: I — Concentration microlocale, Comm. in PDE 19 (1994), no. 9-10, 1535–1563. [3] A. Hassell and T. Tao, Upper and lower bounds for normal derivatives of Dirichlet eigenfunctions, Math. Res. Lett., to appear. [4] B. Helffer, Semi-classical analysis for the Schr¨ odinger operator and applications, Lecture Notes in Mathematics 1336, Springer, Berlin, 1988. [5] L. H¨ ormander, The analysis of linear partial differential operators, volume 3, Springer, Berlin, 1985. [6] S. Ozawa, Asymptotic property of eigenfunction of the Laplacian at the boundary. Osaka J. Math. 30 (1993), 303–314. [7] F. Rellich, Darstellung der Eigenwerte von ∆u + λu durch ein Randintegral. Math. Z., 46, 1940, 635-636. [8] M. Viˇsik and G. Eˇskin, Convolution equations in a bounded region, Russian Mathematical Surveys 20:3 (1965), 86-151. Department of Mathematics, ANU, Canberra, ACT 0200 AUSTRALIA E-mail address: [email protected] Department of Mathematics, UCLA, Los Angeles CA 90095-1555 USA E-mail address: [email protected]
A SYMMETRIC FUNCTIONAL CALCULUS FOR NONCOMMUTING SYSTEMS OF SECTORIAL OPERATORS BRIAN JEFFERIES Abstract. Given a system A = (A1 , . . . , An ) of linear operators whose real linear combinations have spectra contained in a fixed sector in C and satisfy resolvent bounds there, functions f (A) of the system A of operators can be formed for monogenic functions f having decay at zero and infinity in a corresponding sector in Rn+1 . The paper discusses how the functional calculus f 7→ f (A) might be extended to a larger class of monogenic functions and its relationship with a functional calculus for analytic functions in a sector of Cn .
1. Introduction Given a finite system A = (A1 , . . . , An ) of bounded linear operators acting on a Banach space X, it has recently been shown how functions f (A) of the n-tuple A can be formed for a large class of functions f , just under the assumption that the spectrum σ(hA, ξi) of the operator P hA, ξi := nj=1 Aj ξj is a subset of R for every ξ ∈ Rn [5]. The operators A1 , . . . , An do not necessarily commute with each other. A distinguished subset γ(A) of Rn with the property that the bounded linear operator f (A) is defined for any real analytic function f : U → C defined in a neighbourhood U of γ(A) in Rn arises in the approach considered in [5]. For a polynomial p in n real variables, p(A) is the operator formed by substituting symmetric products in the bounded linear operators A1 , . . . , An for the monomial expressions in p, that is, we have a symmetric functional calculus in the n operators A1 , . . . , An . Another way of expressing this symmetry property is that for any ξ ∈ Rn and any polynomial q : C → C in one variable, the equality p(A) = q(hA, ξi) holds for the polynomial p : x 7→ q(hx, ξi), x ∈ Rn . Moreover, the mapping f 7→ f (A) is continuous for a certain topology Date: 3 January 2003. 1991 Mathematics Subject Classification. Primary 47A60, 46H30; Secondary 47A25, 30G35. Key words and phrases. lacunas, Weyl functional calculus, Clifford algebra, monogenic function, symmetric hyperbolic system. 68
NONCOMMUTING SYSTEMS OF SECTORIAL OPERATORS
69
defined on the space of functions real analytic in a neighbourhood of γ(A) in Rn [5, Proposition 3.3]. These properties are analogous to the Riesz-Dunford functional calculus of a single bounded linear operator T acting on X, by which a function f (T ) : X → X of T is defined by the Cauchy integral formula Z 1 (1) f (T ) = (λI − T )−1 f (λ) dλ, 2πi C for f analytic in a neighbourhood of the spectrum σ(T ) of T and for C a simple closed contour about σ(T ). It is by this analogy that the set γ(A) mentioned above may be thought of as the “joint spectrum” of the system A, especially if there is no set smaller than γ(A) possessing the desirable properties alluded to. Now for a single operator T , the spectrum σ(T ) has a simple algebraic definition as the set of all λ ∈ C for which the operator λI − T is not invertible in the space L(X) of bounded linear operators acting on X. In the case that A consists of a system of n commuting, possibly unbounded, linear operators with real spectra, A. McIntosh and A. Pryde [14, 15] gave a simple algebraic definition of the joint spectrum γ(A) of A and used this to obtain operator bounds for solutions of operator equations. Work of A. McIntosh, A. Pryde and W. Ricker [16] established the equivalence of γ(A) with other notions of joint spectrum. In the noncommutative case, we cannot expect such a straightforward algebraic definition of the joint spectrum γ(A), although such a definition was proposed in [8]. Another example of a symmetric functional calculus is the Weyl calculus WA considered in [19] for n selfadjoint operators A = (A1 , . . . , An ). In the case that the system A consists of bounded selfadjoint operators, it was shown in [4] that γ(A) is precisely the support of the operator valued distribution WA , and E. Nelson characterised this set as the Gelfand spectrum of a certain subalgebra of the Banach algebra of operants [17]. Further work along these lines was conducted by E. Albrecht [1]. If we now pass to unbounded operators, then a similar analysis holds if we retain the spectral reality condition σ(hA, ξi) ⊂ R for ξ ∈ Rn , provided that we suitably account for operator domains. However, much of the work [14, 15, 16] on functional calculi just mentioned was motivated by Alan McIntosh’s study of the commuting n-tuple DΣ = (D1 , . . . , Dn ) of differentiation operators on a Lipschitz surface Σ in Rn+1 . In the case that Σ is just the flat surface Rn , the operators Dj = 1i ∂x∂ j , j = 1, . . . , n,
70
BRIAN JEFFERIES
commute with each other and are selfadjoint, otherwise, the unbounded operators Dj , j = 1, . . . , n, have spectra σ(Dj ) contained in a complex sector Sω (C) = {z ∈ C : z 6= 0, | arg(z)| ≤ ω} with an angle ω depending on the variation of the directions normal to the surface Σ. Although the existence and properties of the H ∞ -functional calculus for the commuting n-tuple DΣ are now well-understood, see for example [13], the purpose of the present paper is to initiate a study of the symmetric functional calculus for an n-tuple A of unbounded sectorial operators — we do not assume that the operators commute with each other. In particular, the spectral reality condition (2)
σ(hA, ξi) ⊂ R,
for all ξ ∈ Rn
needs to be relaxed. An alternative approach to forming an H ∞ functional calculus for commuting operators using exponential bounds is given in [11]. Before proceeding with further discussion, we note the definition of the joint spectrum γ(A) for a system A satisfying condition (2). The key idea behind [14, 15] in the commuting case and [4, 5, 8, 9] in the noncommuting case, is to produce a higher-dimensional analogue of the Riesz-Dunford formula (1). So what we need is a higher-dimensional analogue of the Cauchy integral formula in complex analysis and then, in the time-honoured fashion of operator theory, substitute an n-tuple of numbers by an n-tuple of unbounded linear operators. But this is easier said than done. It turns out that Clifford analysis provides a higher dimensional analogue of the Cauchy integral formula especially well-suited to the noncommutative setting. Even for the commuting n-tuple DΣ of operators mentioned above, it provides the connection between multiplier operators and singular convolution operators for functions defined on a Lipschitz surface. A brief r´esum´e of Clifford analysis [2, 3] and the monogenic functional calculus treated in [5] follows. Let C(n) be the Clifford algebra generated over the field C by the standard basis vectors e0 , e1 , . . . , en of Rn+1 with conjugation u 7→ u. P The generalized Cauchy-Riemann operator is given by D = nj=0 ej ∂x∂ j . Let U ⊂ Rn+1 be an open set. A function f : U → C(n) is called left monogenic if Df = 0 in U and right monogenic if f D = 0 in U . The Cauchy kernel is given by (3)
Gx (y) =
1 x−y , σn |x − y|n+1
x, y ∈ Rn+1 , x 6= y,
NONCOMMUTING SYSTEMS OF SECTORIAL OPERATORS
71
n+1 with σn = 2π 2 /Γ n+1 the volume of unit n-sphere in Rn+1 . So, 2 given a left monogenic function f : U → C(n) defined in an open subset U of Rn+1 and an open subset Ω of U such that the closure Ω of Ω is contained in U , and the boundary ∂Ω of Ω is a smooth oriented n-manifold, then the Cauchy integral formula Z f (y) = Gx (y)n(x)f (x) dµ(x), y ∈ Ω ∂Ω
is valid. Here n(x) is the outward unit normal at x ∈ ∂Ω and µ is the volume measure of the oriented manifold ∂Ω. An element x = (x0 , x1 , . . . , xn ) of Rn+1 will often be written as x = x0 e0 + ~x with P ~x = nj=1 xj ej . By analogy with formula (1), our aim is to define Z (4) f (A) = Gx (A)n(x)f (x) dµ(x) ∂Ω
for the n-tuple A = (A1 , . . . , An ) of bounded linear operators on X. A difficulty occurs in making sense of the Cauchy kernel x 7→ Gx (A), a function with values in the space L(X) ⊗ C(n) that should be defined and two-sided monogenic for all x off a nonempty closed subset γ(A) of Rn inside Ω. The set ∂Ω can be smoothly varied in the region where x 7→ Gx (A) is right-monogenic. Of course, one would also like f (A) to be the ‘correct’ operator in the case that f is the unique monogenic extension to Rn+1 of a polynomial in n variables. In the Riesz-Dunford functional calculus for T , the set of singularities of the resolvent λ 7→ (λI − T )−1 is precisely the spectrum σ(T ) of T , so the set γ(A) may be interpreted as a higher-dimensional analogue of the spectrum of a single operator. It seems reasonable to call the set γ(A) the monogenic spectrum of the n-tuple A by analogy with the case of a single operator. The program was implemented by A. McIntosh and A. Pryde for commuting n-tuples of bounded operators with real spectrum in order to give estimates on the solution of systems of operator equations [14, 15]. In the case that n is odd, we have ( )c n X γ(A) = λ ∈ Rn : (λj I − Aj )2 is invertible in L(X) j=1
and n
X 1 Gx (A) = (x − A) x20 I + (xj I − Aj )2 σn j=1
!− n+1 2 ,
72
BRIAN JEFFERIES
for all x = (x0 , . . . , xn ) ∈ Rn+1 \ ({0} × γ(A)). It turns out that γ(A) coincides with the Taylor spectrum for commuting systems of bounded linear operators [16]. If the n-tuple A of bounded linear operators satisfies exponential growth conditions, such as when A1 , . . . , An are selfadjoint, then Weyl’s functional calculus WA is associated with A and Gx (A) = W(Gx ) is an obvious way to define the Cauchy kernel for all x outside the support of WA . It is shown in [4] that formula (4) holds. However, in this case, we actually have a symmetric functional calculus defined over γ(A) richer than just all real analytic functions. The Cauchy kernel Gx (A) can also be written as a series expansion like the Neuman series for the resolvent of a single operator if x ∈ Rn+1 lies outside a sufficiently large ball [8, 9], but the expansion does not allow us to identify γ(A) as a subset of Rn in the case that the spectral reality condition (2) holds. A third way to define the Cauchy kernel Gx (A) for the monogenic functional calculus whenever the spectral reality condition (2) holds, is by the plane wave decomposition for the Cauchy kernel (3) given by F. Sommen [18]. This was investigated by A. McIntosh and J. PictonWarlow soon after the papers [14, 15] appeared. The formula is n (n − 1)! i Gx (A) = sgn(x0 )n−1 2 2π (5) Z (e0 + is) (h~x, siI − hA, si − x0 sI)−n ds × S n−1
for all x = x0 e0 + ~x with x0 a nonzero real number and ~x ∈ Rn . Here S n−1 is the unit (n − 1)-sphere in Rn , ds is surface measure and the inverse power (h~xI − A, si − x0 s)−n is taken in the Clifford module L(X)⊗C(n) . The spectral reality condition (2) ensures the invertibility of (h~xI − A, si − x0 s) for all x0 6= 0 and s ∈ S n−1 by the spectral mapping theorem. Even if A satisfies exponential growth conditions, with the left hand side given by formula (5), the equality Gx (A) = WA (Gx ) can still be used to good effect. In [6], it was used to geometrically characterise the support of fundamental solution of the symmetric hyperbolic system associated with a pair A of hermitian matrices in the case n = 2. Greater understanding of Clifford residue theory would enable a similar treatment in higher dimensions. In Section 2, it is shown how formula (5) for the Cauchy kernel associated with the system A of sectorial operators still works if the
NONCOMMUTING SYSTEMS OF SECTORIAL OPERATORS
73
spectral reality condition (2) is replaced by a sectoriality condition with the appropriate resolvent bounds. The system DΣ of commuting sectorial operators described above is of this type. By this means functions f (A) of the operators A can be formed, provided that f is, say, left monogenic in a sector in Rn+1 and satisfies suitable decay estimates at 0 and ∞, in a fashion similar to the case of a single operator of type ω [12]. Because Gx (A) is only defined for x outside a sector in Rn+1 , the monogenic spectrum γ(A) is now contained in that sector in Rn+1 . Recall that under condition (2), γ(A) is a subset of Rn . A function f (DΣ ) of the system DΣ has a natural interpretation as a multiplier operator acting on Lp -spaces of functions defined on the Lipschitz surface Σ, as well as a singular convolution operator, so the multiplier f should be a bounded analytic function defined on a sector in Cn [13], rather than, say, a bounded monogenic function defined in a sector in Rn+1 . The monogenic functional calculus for a system A of sectorial operators appears to be moving us inexorably in the wrong direction. The problem arises of establishing a bijection between monogenic functions defined on a sector in Rn+1 and analytic functions defined on a sector in Cn , together with the appropriate norms — this is a question of function theory rather than operator theory. The association is via the Cauchy-Kowaleski extension to a sector in Rn+1 of the restriction of the analytic function to Rn \ {0}. The purpose of this paper is to make some observations about the relationship between monogenic functions defined in a sector in Rn+1 and analytic functions defined in the corresponding sector in Cn , with applications to functional calculi of systems of operators firmly in mind. In Section 3, the spectral properties of multiplication operators in C(n) are examined along the lines of Lecture 1 of [13]. In Section 4, this enables us to uniquely associate a bounded analytic function defined in a sector in Cn with a suitably decaying monogenic function defined in the corresponding sector in Rn+1 via the Cauchy integral formula. 2. The plane wave decomposition Let A = (A1 , . . . , An ) be an n-tuple of densely defined linear operators Aj : D(Aj ) → X acting in X such that ∩nj=1 D(Aj ) is dense in X. The space L(n) (X(n) ) of left module homomorphisms of X(n) = X ⊗C(n) is identified with L(X) ⊗ C(n) in the natural way and becomes a right Banach module under the uniform operator topology.
74
BRIAN JEFFERIES
If we take formula (5) as the definition of Gx (A), then the convergence of the integral Z
(e0 + is) (h~xI − A, si − x0 sI)−n ds
S n−1
for particular values of x = x0 e0 + ~x ∈ Rn+1 is at issue. Now (h~xI − A, si − x0 sI)−1 = (h~xI − A, si + x0 sI) h~xI − A, si2 + x20 I
−1
if 0 ∈ / σ (h~xI − A, si2 + x20 ). Thus, we need to ensure the appropriate uniform operator bounds for −1 h~xI − A, si2 + x20 I , s ∈ S n−1 as x = x0 e0 + x ranges over a subset of Rn+1 . In the case that σ(hA, si) ⊂ R and (λI − hA, si)−1 is suitably bounded for all s ∈ S n−1 and λ ∈ C \ R, then Gx0 e0 +~x (A) is defined for all x0 6= 0. First, for each 0 < ν < π/2, set Sν+ (C) = {z ∈ C : | arg z| ≤ ν} ∪ {0}, Sν (C) = Sν+ (C) ∪ ig(−Sν+ (C) , ¯ ◦ Sν+ (C) = {z ∈ C : | arg z| < ν}, ◦ ◦ Sν◦ (C) = Sν+ (C) ∪ − Sν+ (C) . The (n − 1)-sphere in Rn is denoted by S n−1 . Definition 2.1. Let A = (A1 , . . . , An ) be an n-tuple of densely defined T linear operators Aj : D(Aj ) → X acting in X such that nj=1 D(Aj ) is dense in X and let 0 ≤ ω < π2 . Then A is said to be uniformly of type ω if for every s ∈ S n−1 , the operator hA, si is closable with closure hA, si, the inclusion σ hA, si ⊂ Sω (C) holds, and for each ν > ω, there exists Cν > 0 such that −1 (6) k zI − hA, si k ≤ Cν |z|−1 , z ∈ / Sν◦ (C), s ∈ S n−1 . It follows that s 7→ hA, si is continuous on S n−1 in the sense of strong −1 resolvent convergence [7, Theorem VIII.1.5]. Because zI − hA, si is densely defined and uniformly bounded in X, the closure symbol will be omitted. Now suppose that equation (6) is satisfied and let z = h~x, si + ix0 . Then z ∈ / Sν◦ (C) means that | arg z| ≥ ν for − π2 ≤ arg z ≤ π2 or π − arg z ≥ ν for π2 ≤ arg z ≤ π or π + arg z ≥ ν for −π ≤ arg z ≤ − π2 . Hence, we have |x0 | ≥ tan ν|h~x, si|.
NONCOMMUTING SYSTEMS OF SECTORIAL OPERATORS
75
First, let Nν = {x ∈ Rn+1 : x = x0 e0 + ~x, |x0 | ≥ tan ν|~x| }, Sν (Rn+1 ) = {x ∈ Rn+1 : x = x0 e0 + ~x, |x0 | ≤ tan ν|~x| }, Sν◦ (Rn+1 ) = {x ∈ Rn+1 : x = x0 e0 + ~x, |x0 | < tan ν|~x| }. Note that if x0 e0 + x ∈ Nν , then z = h~x, si + ix0 ∈ / Sν◦ (C) for every s ∈ S n−1 , because either |x0 | ≥ tan ν|~x| ≥ tan ν|h~x, si|. Lemma 2.2. Let ω < ν < π/2. Suppose that the n-tuple A of linear operators is uniformly of type ω. Then for all x0 e0 + ~x ∈ Nν , the integral Z
(h~xI − A, si − x0 sI)−n ds L(n) (X(n) )
S n−1
converges and satisfies the bound Z
(h~xI − A, si − x0 sI)−n L
(n) (X(n)
S n−1
ds ≤ )
Cν0 . |x0 |n
Proof. For every x0 e0 + ~x ∈ Nν , we have z = hx, si ± ix0 ∈ / Sν (C) so that the operator (h~x, si ± ix0 )I − hA, si is invertible and the bound
Cν
((h~x, si ± ix0 )I − hA, si)−1 p ≤ L(X) h~x, si2 + x20 holds. Now (h~xI − A, si − x0 sI)−1 = (h~xI − A, si + x0 sI) h~xI − A, si2 + x20 I
−1
where h~xI − A, si2 + x20 I
−1
= ((h~x, si + ix0 )I − hA, si)−1 ((h~x, si − ix0 )I − hA, si)−1 . Writing (h~xI − A, si + x0 sI) = ((h~x, si + ix0 )I − hA, si) − ix0 I + x0 sI, we obtain (h~xI − A, si − x0 sI)−1 = ((h~x, si − ix0 )I − hA, si)−1 − ix0 (e0 + is) h~xI − A, si2 + x20 I so that by the estimate (6) we have
(h~xI − A, si − x0 sI)−1 ≤ p L (X ) (n)
(n)
Cν
h~x, si2 + x20 Cν + 2Cν2 ≤ , |x0 |
from which the stated bound follows.
+
−1
,
2|x0 |Cν2 h~x, si2 + x20
76
BRIAN JEFFERIES
Thus, if A is uniformly of type ω, then x0 e0 + ~x 7→ Gx0 e0 +~x (A) is defined by equation (5) for all x0 e0 + ~x ∈ Nν with ω < ν < π/2. Standard arguments ensure that x0 e0 + ~x 7→ Gx0 e0 +~x (A) is both left and right monogenic as an element of L(X) ⊗ C(n) . If we denote by γ(A) ⊂ Rn+1 the set of all singularities of the function x0 e0 + ~x 7→ Gx0 e0 +~x (A), then γ(A) ⊆ Sω (Rn+1 ). Suppose that ω < ν < π/2, 0 < s < n and f is a left monogenic function defined on Sν◦ (Rn+1 ) such that for every 0 < θ < ν there exists Cθ > 0 such that |x|s (7) |f (x)| ≤ Cθ , x ∈ Sθ◦ (Rn+1 ). (1 + |x|2s ) According to Lemma 2.2, for every ω < ν 0 < θ < ν, we have |x|s kGx (A)k.|f (x)| ≤ Cθ,ν 0 , x = x0 e0 + ~x |x0 |n (1 + |x|2s ) for all x ∈ Sθ◦ (Rn+1 ) ∩ Nν 0 . Now if ω < θ < ν and (8)
Hθ = {x ∈ Rn+1 : x = x0 e0 + ~x, |x0 |/|x| = tan θ} ⊂ Sν◦ (Rn+1 ).
it follows that kGx (A)k.|f (x)| = O(1/|x|n−s ) as x → 0 in Hθ . Hence, x 7→ Gx (A)n(x)f (x) is locally integrable at zero with respect to ndimensional surface measure on Hθ . Similarly, kGx (A)k.|f (x)| = O(1/|x|n+s ) as |x| → ∞ in Hθ , so x 7→ Gx (A)n(x)f (x) is integrable with respect to n-dimensional surface measure on Hθ . Therefore, we define Z (9) f (A) = Gx (A)n(x)f (x) dµ(x). Hθ
If ψ : R \ {0} → C has a two-sided monogenic extension ψ˜ to ◦ n+1 ˜ Sν (R ) that satisfies the bound (7) for all 0 < θ < ν, then ψ(A) is written just as ψ(A). Formula (9) does just what we would expect in the noncommuting situation. For example, let p be a polynomial of degree n with p(0) = 0 and bλ (z) = p(z)(λ − z)−n−1 for some λ ∈ / Sν◦ (C). Let ξ ∈ Rn and set φλ,ξ (x) = bλ (hx, ξi) for all x ∈ Rn . Denote the two-sided monogenic extension of φλ,ξ to Sν◦ (Rn+1 ) by φ˜λ,ξ . Then φ˜λ,ξ has decay at zero and infinity and we have φλ,ξ (A) = φ˜λ,ξ (A) = p(hA, ξi)(λI − hA, ξi)−n−1 is a bounded linear operator. In order to form functions f (A) of the system A of operators for a class of monogenic functions f larger than those which satisfy a n+1
NONCOMMUTING SYSTEMS OF SECTORIAL OPERATORS
77
bound like (7), a greater understanding of function theory in the sector Sω (Rn+1 ) is needed. To this end, the simple system A = ζ = (ζ1 , . . . , ζn ) ∈ Cn of multiplication operators in the algebra C(n) is considered in the next section. 3. Joint spectral theory in the algebra C(n) Let ζ = (ζ1 , . . . , ζn ) be a vector belonging to Cn . The complex spectrum σ(iζ) of the element iζ = i(ζ1 e1 + · · · + ζn en ) of the algebra C(n) is σ(iζ) = {λ ∈ C : (λe0 − iζ) does not have an inverse in C(n) }. Following [13, Section 5.2], we check that (λe0 + iζ)(λe0 − iζ) = λ2 e0 − i2 ζ 2 = (λ2 − |ζ|2C )e0 , P where |ζ|2C = nj=1 ζj2 . So, for all λ ∈ C for which, λ 6= ±|ζ|C , the element (λe0 − iζ) of the algebra C(n) is invertible and (λe0 − iζ)−1 =
λe0 + iζ . λ2 − |ζ|2C
If |ζ|2C 6= 0, the two square roots of |ζ|2C are written as ±|ζ|C and |ζ|C = 0 for |ζ|2C = 0. Hence σ(iζ) = {±|ζ|C }. When |ζ|2C 6= 0, the spectral projections 1 iζ χ± (ζ) = e0 + 2 ±|ζ|C are associated with each element ±|ζ|C of the spectrum σ(iζ) and iζ has the spectral representation iζ = |ζ|C χ+ (ζ) + (−|ζ|C )χ− (ζ). Henceforth, we use the symbol |ζ|C to denote the positive square root of |ζ|2C in the case that |ζ|2C ∈ / (− ∈ f ty, 0]. On the other hand, according to the point of view mentioned in the Introduction, the monogenic spectrum γ(ζ) of ζ ∈ Cn should be the set of singularities of the Cauchy kernel x 7→ Gx (ζ) in the algebra C(n) . Although Gx (ζ) is defined by formula (3) only for ζ ∈ Rn and x 6= ζ, a natural choice for the Cauchy kernel for ζ ∈ Cn is to take the maximal analytic extension ζ 7→ Gx (ζ) of formula (3) for ζ ∈ Cn , that is, (10) 1 x+ζ |x − ζ|2C ∈ / (−∞, 0], n even n+1 Gx (ζ) = , x ∈ R , 2 n+1 |x − ζ|C 6= 0, n odd σn |x − ζ|C P Here |x − ζ|2C = x20 + nj=1 (xj − ζj )2 and |x − ζ|C is the positive square root of |x − ζ|2C , coinciding with the analytic extension of the modulus function ξ 7→ |x − ξ|, ξ ∈ Rn \ {x}.
78
BRIAN JEFFERIES
The analogous reasoning for multiplication by x ∈ Rn+1 in the algebra C(n) just gives us the Cauchy kernel (3), so that γ(x) = {x}, as expected. Remark 3.1. If ζ = (ζ1 , . . . , ζn ) satisfies the conditions of Definition 2.1, then there exists θ ∈ [−ω, ω] and x ∈ Rn such that ζ = eiθ x. To see this, write ζ = α + iβ for α, β ∈ Rn . If |hβ, ξi| ≤ |hα, ξi| tan ω for all ξ ∈ Rn , then α⊥ ⊂ β ⊥ , so that β ∈ span{α}. In this case, the plane wave formula (5) with A = ζ and equation (10) agree by analytic continuation, at least for x ∈ Nν with ν > |θ|. Given ζ ∈ Cn , if singularities of (10) occur at x ∈ Rn+1 , then |x − ζ|2C ∈ (−∞, 0], otherwise we can simply take the positive square root of |x − ζ|2C in formula (10) to obtain a monogenic function of x. To determine this set, write ζ = ξ + iη for ξ, η ∈ Rn and x = x0 e0 + ~x for x0 ∈ R and ~x ∈ Rn . Then |x − ζ|2C = x20 +
n X
(xj − ζj )2
j=1
= x20 +
n X
(xj − ξj − iηj )2
j=1
(11)
=
x20
+ |~x − ξ|2 − |η|2 − 2ih~x − ξ, ηi.
Thus, |x−ζ|2C belongs to (−∞, 0] if and only if x lies in the intersection of the hyperplane h~x − ξ, ηi = 0 passing through ξ and with normal η, and the ball x20 + |~x − ξ|2 ≤ |η|2 centred at ξ with radius |η|. If n is even, then (12) γ(ζ) = {x = x0 e0 +~x ∈ Rn+1 : h~x−ξ, ηi = 0, x20 +|~x−ξ|2 ≤ |η|2 }. and if n is odd, then (13) γ(ζ) = {x = x0 e0 +~x ∈ Rn+1 : h~x−ξ, ηi = 0, x20 +|~x−ξ|2 = |η|2 }. In particular, if =(ζ) = 0, then γ(ζ) = {ζ} ⊂ Rn . Remark 3.2. The distinction between n odd and even is reminiscent of the support of the fundamental solution of the wave equation in even and odd dimensions. Off γ(ζ), the function x 7→ Gx (ζ) is clearly two-sided monogenic, so the Cauchy integral formula gives Z Gx (ζ)n(x)f (x) dµ(x) (14) f˜(ζ) = ∂Ω
NONCOMMUTING SYSTEMS OF SECTORIAL OPERATORS
79
for a bounded open neighbourhood Ω of γ(ζ) with smooth oriented boundary ∂Ω, outward unit normal n(x) at x ∈ ∂Ω and surface measure µ. The function f is assumed to be left monogenic in a neighbourhood of Ω, but ζ 7→ f˜(ζ) is an analytic C(n) -valued function as the closed set γ(ζ) varies inside Ω. Moreover, f˜ equals f on Ω ∩ Rn by the usual Cauchy integral formula of Clifford analysis, so if f is, say, the monogenic extension of a polynomial p : Cn → C restricted to Rn , then f˜(ζ) = p(ζ), as expected. In this way, for each left monogenic function f defined in a neighbourhood of γ(ζ), in a natural way we associate an analytic function f˜ defined in a neighbourhood of ζ. It is clear that if ζ = ξ + iη lies in a sector in Cn , say, |η| ≤ |ξ| tan ν, then the monogenic spectrum γ(ζ) lies in a corresponding sector in Rn+1 . More precisely, we have Proposition 3.3. Let ζ ∈ Cn \ {0} and 0 < ω < π/2. Then γ(ζ) ⊂ Sω (Rn+1 ) if and only if (15)
|ζ|2C 6= (−∞, 0]
and
|=(ζ)| ≤ <(|ζ|C ) tan ω.
Proof. The statement is trivially valid if ζ ∈ Rn \ {0}, so suppose that =(ζ) 6= 0. Then the monogenic spectrum γ(ζ) of ζ given by (12) is a subset of Sω (Rn+1 ) if and only if there exists 0 < θ ≤ ω such that the cone Hθ+ = {x0 e0 + ~x ∈ Rn+1 : x0 > 0, x0 = |x| tan θ } is tangential to the boundary of γ(ζ). A calculation shows that Hθ+ is tangential to the boundary of γ(ζ) for all ζ = ξ + iη with ξ, η ∈ Rn , satisfying (16)
|η|2 = sin2 θ(|ξ|2 + tan2 θ|Pη ξ|2 ).
Here Pη : u 7→ hu, ηiη/|η|2 , u ∈ Rn , is the projection operator onto span{η}. To relate condition (16) to the inequality (15), suppose that m = m0 e0 + m ~ is the unit vector normal to Hθ such that m ~ lies in the direction of η. Hence, m0 = cot θ|m|, ~ tan θ = |m|/m ~ 0 and Pη ξ = 2 hξ, mi ~ m/| ~ m| ~ . Then equation (15) becomes η = sin θ(m20 |ξ|2 + hξ, mi ~ 2 )1/2
m ~ . |m|m ~ 0
But |m0 e0 + m| ~ = 1, so (cot2 θ + 1)|m| ~ 2 = 1. We have |m| ~ = sin θ and (17)
η = (m20 |ξ|2 + hξ, mi ~ 2 )1/2
m ~ . m0
80
BRIAN JEFFERIES
As mentioned in [13, p67], the set of all ζ = ξ + iη with η 6= 0 satisfying (17) is equal to the set of all ζ = ξ + iη with η 6= 0 satisfying m ~ |ζ|2C 6= (−∞, 0] and η = <(|ζ|C ) . m0 Because |m|/m ~ 0 = tan θ ≤ tan ω, we obtain the desired equivalence by letting m ~ vary over all directions in Rn . For each 0 < ω < π/2, let Sω (Cn ) denote the set of all ζ ∈ Cn satisfying condition (8) and let Sω◦ (Cn ) be its interior. Corollary 3.4. Let f : Sω◦ (Rn+1 ) → C(n) be a left monogenic function such that the restriction f˜ of f to Rn \ {0} takes values in C. Then f˜ is the restriction to Rn \ {0} of an analytic function defined on Sω◦ (Cn ) The sectors Sω (Cn ) ⊂ Cn and Sω (Rn+1 ) ⊂ Rn+1 are dual to each other in the sense that the mapping (ω, ζ) 7→ Gω (ζ),
ω ∈ Rn+1 \ Sω (Rn+1 ), ζ ∈ Sω◦ (Cn )
is two-sided monogenic in ω and analytic in ζ. The sector Sω (Cn ) arose in [10] as the set of ζ ∈ Cn for which the exponential functions e+ (x, ζ) = eih~x,ζi e−x0 |ζ|C χ+ (ζ),
x = x0 e0 + ~x,
have decay at infinity for all x ∈ Rn+1 with hx, mi > 0 and all unit vectors m = m0 e0 + m ~ ∈ Rn+1 satisfying m0 ≥ cot ω|m|. ~ 4. Joint spectral theory of sectorial multiplication operators By means of the higher-dimensional analogue(4)of the Riesz-Dunford functional calculus, we can form functions f (A) of a noncommuting system A of operators uniformly of type ω for left monogenic functions f defined on a sector Sν (Rn+1 ), ω < ν < π/2, provided that f has decay at zero and infinity. The observations of the preceding section mean that we can do something similar for the commutative system of multiplication operators in the sector Sω (Cn ), although these are not uniformly of type ω. The problem mentioned in the Introduction of connecting monogenic functions defined on a sector in Rn+1 with analytic functions defined on a sector in Cn can be reformulated simply in terms of studying the monogenic functional calculus for multiplication operators. More precisely, let 0 < ω < π/2 and set X = L2 (Sω (Cn ), C(n) ). Integration is with respect to Lebesgue measure on Cn ≡ R2n . Let
NONCOMMUTING SYSTEMS OF SECTORIAL OPERATORS
81
Mj be the operator of multiplication by the j’th coordinate function defined on Sω (Cn ), that is, the domain D(Mj ) of Mj is the set of all functions ψ ∈ X such that the function Mj ψ defined by (18)
(Mj ψ)(ζ) = ζj ψ(ζ),
ζ ∈ Sω (Cn ),
belongs to X, and the unbounded operator Mj is given by formula (18) for each ψ ∈ D(Mj ). Set M = (M1 , . . . , Mn ). Then M is a commuting n-tuple of normal operators. Unlike the n-tuple DΣ of operators mentioned in the Introduction, the existence of a joint functional calculus for M is not an issue. Indeed, the joint spectral measure P : B(Sω (Cn )) → L(X) given by P (B)ψ = χB .ψ,
ψ ∈ X, B ∈ B(Sω (Cn )),
has support Sω (Cn ). For any bounded Borel measurable function f : Sω (Cn ) → C, the bounded linear operator Z f (M ) = f dP Sω (Cn )
is given by the functional calculus for commuting normal operators, and explicitly, by f (M ) : ψ 7→ f.ψ,
ψ ∈ X.
Thus we have a functional calculus for M for a class of functions far richer than uniformly bounded analytic functions f defined in a sector Sν◦ (Cn ) with ω < ν < π/2. Nevertheless, it is not so obvious that bounded linear operators f (M ) can also be formed naturally for functions f that are monogenic in a sector Sν◦ (Rn+1 ) with ω < ν < π/2. The Cauchy kernel Gx (M ) is defined for all x ∈ Nν and all ν such that ω < ν < π/2 simply by setting (19)
(Gx (M )ψ)(ζ) = ψ(ζ)Gx (ζ),
ζ ∈ Sω (Cn ),
for all ψ ∈ X and x ∈ Nν , so that Gx (M ) is a left module homomorphism of X. The proof of the next lemma is straightforward and is omitted. Lemma 4.1. Let ω < ν < π/2. Then Gx (M ) is a bounded linear operator on X for all x ∈ Nν \{0}. Furthermore, the operator valued function x 7→ Gx (M ), x ∈ Nν \ {0}, is continuous for the strong operator topology, right monogenic in the interior of Nν and kGx (M )k = O(|x|−n ) as |x| → ∞ and |x| → 0 in Nν
82
BRIAN JEFFERIES
Let ω < ν < π/2. Suppose that f is a left monogenic function defined on Sν◦ (Rn+1 ) such that for every 0 < θ < ν, the bound (7) holds. Now if ω < θ < ν a nd Hθ is the two-sheeted cone (8) in Rn+1 , then the bounded linear operator f (M ) : X → X is defined by the formula Z (20) f (M ) = Gx (M )n(x)f (x) dµ(x) Hθ
by the same argument by which (9) is defined. Then f (M ) is a left module homomorphism of X. The operator f (M ) is identified in the next statement. Proposition 4.2. Suppose that f is a left monogenic function defined on Sν◦ (Rn+1 ) satisfying the bound (7) for every 0 < θ < ν and f (M ) is defined by formula (20). Let f˜ be the C(n) -valued analytic function defined from f by formula (14). Then f˜ is uniformly bounded on the sector Sω (Cn ), (21)
(f (M )ψ)(ζ) = ψ(ζ)f˜(ζ),
ζ ∈ Sω (Cn ),
for all ψ ∈ X and kf (M )k = kf˜kH ∞ (Sω (Cn )) . The problem remains of obtaining better bounds for the operator norm of f (M ) in terms of bounds of the left monogenic function defined on Sν◦ (Rn+1 ) and feeding these bounds back into formula (9) for a system A uniformly of type ω. References [1] E. Albrecht, Several variable spectral theory in the non-commutative case, Spectral Theory, Banach Center Publications, Volume 8, PWN-Polish Scientific Publishers, Warsaw, 1982, 9–30. [2] F. Brackx, R. Delanghe and F. Sommen, Clifford Analysis, Research Notes in Mathematics 76, Pitman, Boston/London/Melbourne, 1982. [3] R. Delanghe, F. Sommen and V. Soucek, Clifford Algebras and Spinor Valued Functions: A Function Theory for the Dirac Operator, Kluwer, Dordrecht, 1992. [4] B. Jefferies and A. McIntosh, The Weyl calculus and Clifford analysis, Bull. Austral. Math. Soc. 57 (1998), 329–341. [5] B. Jefferies, A. McIntosh and J. Picton-Warlow, The monogenic functional calculus, Studia Math. 136 (1999), 99–119. [6] B. Jefferies and B. Straub, Lacunas in the support of the Weyl calculus for two hermitian matrices, J. Austral. Math. Soc. (Series A), to appear. [7] T. Kato, Perturbation Theory for Linear Operators, 2nd Ed., Springer-Verlag, Berlin/Heidelberg/New York, 1980. [8] V. V. Kisil, M¨ obius transformations and monogenic functional calculus, ERA Amer. Math. Soc. 2 (1996), 26–33
NONCOMMUTING SYSTEMS OF SECTORIAL OPERATORS
83
[9] V. V. Kisil and E. Ram´ırez de Arellano, The Riesz-Clifford functional calculus for non-commuting operators and quantum field theory, Math. Methods Appl. Sci. 19 (1996), 593–605. [10] C. Li, A. McIntosh, T. Qian, Clifford algebras, Fourier transforms and singular convolution operators on Lipschitz surfaces, Rev. Mat. Iberoamericana 10 (1994), 665–721. [11] C. Li, A. McIntosh, Clifford algebras and H∞ functional calculi of commuting operators, Clifford algebras in analysis and related topics (Fayetteville, AR, 1993), 89–101, Stud. Adv. Math., CRC, Boca Raton, FL, 1996. [12] A. McIntosh, Operators which have an H∞ -functional calculus, in: Miniconference on Operator Theory and Partial Differential Equations 1986, 212–222 Proc. Centre for Mathematical Analysis 14, ANU, Canberra, 1986. [13] , Clifford algebras, Fourier theory, singular integrals, and harmonic functions on Lipschitz domains, Clifford algebras in analysis and related topics (Fayetteville, AR, 1993), 33–87, Stud. Adv. Math., CRC, Boca Raton, FL, 1996. [14] A. McIntosh and A. Pryde, The solution of systems of operator equations using Clifford algebras, in: Miniconference on Linear Analysis and Function Spaces 1984, 212–222, Proc. Centre for Mathematical Analysis 9, ANU, Canberra, 1985. , A functional calculus for several commuting operators, Indiana U. [15] Math. J. 36 (1987), 421–439. [16] A. McIntosh, A. Pryde and W. Ricker, Comparison of joint spectra for certain classes of commuting operators, Studia Math. 88 (1988), 23–36. [17] E. Nelson, Operants: A functional calculus for non-commuting operators, in: Functional analysis and related fields, Proceedings of a conference in honour of Professor Marshal Stone, Univ. of Chicago, May 1968 (F.E. Browder, ed.), Springer-Verlag, Berlin/Heidelberg/New York, 1970, pp. 172–187. [18] F. Sommen, Plane wave decompositions of monogenic functions, Annales Pol. Math. 49 (1988), 101–114. [19] M.E. Taylor, Functions of several self-adjoint operators, Proc. Amer. Math. Soc. 19 (1968), 91–98. School of Mathematics, The University of New South Wales, Sydney, NSW 2052, Australia E-mail address: [email protected]
SIMILARITIES OF
ω -ACCRETIVE
OPERATORS
CHRISTIAN LE MERDY Abstract. Given a number 0 < ω ≤ π2 , an ω-accretive operator is a sectorial operator A on Hilbert space whose numerical range lies in the closed sector of all z ∈ C such that |Arg(z)| ≤ ω. It is easy to check that any such operator admits bounded imaginary powers, with kAit k ≤ eω|t| for any t ∈ R. We show that conversely, A is similar to an ω-accretive operator if kAit k ≤ eω|t| for any t ∈ R.
1. Introduction. Let H be a Hilbert space and let A be a closed operator on H with dense domain D(A). Given any ω ∈ (0, π), we let Σω be the open sector of all complex numbers z ∈ C∗ such that |Arg(z)| < ω, and we say that A is sectorial of type ω if its spectrum σ(A) is included in the closure of Σω and if for every θ ∈ (ω, π), the set z(z − A)−1 : z ∈ / Σθ is bounded. Assume that ω ≤ π2 . We say that A is ω-accretive if it is sectorial of type ω and if (1.1)
ξ ∈ D(A).
hAξ, ξi ∈ Σω ,
It is well-known that if the resolvent set ρ(A) contains −1, say, then (1.1) implies that A is sectorial of type ω. Thus A is ω-accretive if and only if −1 ∈ ρ(A) and (1.1) holds true. Note that with this terminology, π -accretivity coincides with maximal accretivity. The aim of this note 2 is to give a characterization of injective ω-accretive operators up to similarity in terms of their imaginary powers. If A is an injective maximal accretive operator on H, then we can π define its imaginary powers and we have kAit k ≤ e 2 |t| for any real number t ∈ R. Indeed this estimate is a consequence of von Neumann’s inequality, see e.g. [1, Theorem G]. More generally, assume that A is an π π injective ω-accretive operator. Then ei( 2 −ω) A and e−i( 2 −ω) A are both π π maximal accretive hence for any t ∈ R, we have k(ei( 2 −ω) A)it k ≤ e 2 |t| π π and k(e−i( 2 −ω)A )it k ≤ e 2 |t| . We easily deduce that (1.2)
kAit k ≤ eω|t| , 84
t ∈ R.
SIMILARITIES OF
ω -ACCRETIVE
OPERATORS
85
Our main result asserts that conversely, if A is an injective sectorial operator satisfying (1.2), then A is similar to an ω-accretive operator, that is, there exists a bounded and invertible operator S : H → H such that S −1 AS is ω-accretive. We thus have the following characterization. Theorem 1.1. Let ω ∈ (0, π2 ] be a number and let A be an injective sectorial operator on H. Then A is similar to an ω-accretive operator if and only if there exists a bounded and invertible operator S : H → H such that kS −1 Ait Sk ≤ eω|t| for any t ∈ R. We wish to make three comments concerning this theorem. First, it complements a previous result of ours ([7]) saying that if A is an injective sectorial operator of type < π2 , then A is similar to a maximal accretive operator if and only if it admits bounded imaginary powers. Second, Simard’s recent work ([12]) shows that our result is essentially optimal. Indeed on the one hand, [12, Theorem 1] implies that for any ω ≤ π2 , one can find A not similar to an ω-accretive operator whose imaginary powers satisfy an estimate kAit k ≤ Keω|t| for some K > 1. On the other hand, [12, Theorem 4] shows that one can find A π satisfying kAit k ≤ e 2 |t| for any t ∈ R without being maximal accretive. The third comment is that our proof heavily relies on some recent work of Crouzeix and Delyon ([5]) who established some remarkable estimates for the analytic functional calculus associated to an operator whose numerical range lies in a band of the complex plane. We now give a consequence of Theorem 1.1 concerning fractional powers of ω-accretive operators. Let 0 < ω ≤ π2 and α ∈ (0, 1] be two numbers. It is well-known that if A is an ω-accretive operator, then Aα is αω-accretive. Although the converse does not hold true (see e.g. the discussion at the end of [12]), Theorem 1.1 implies the following. Corollary 1.2. Let A be an ω-accretive operator for some ω ≤ π2 and 1 let α ≥ 2ω be a number. Then A α is similar to an ωα -accretive operator. π Proof. We may assume that A is injective and that α ≤ 1. Then our 1 t assumption of ω-accretivity implies (1.2). Since (A α )it = Ai α , we thus 1 ω have k(A α )it k ≤ e α |t| for any t ∈ R. According to Theorem 1.1, this 1 implies that A α is similar to an ωα -accretive operator, whence the result by taking α-th powers. The proof of Theorem 1.1 is given in Section 3. It uses both H ∞ functional calculus techniques (as introduced by McIntosh in [8]) and a theorem of Paulsen ([9]) reducing our proof to the study of the complete
86
CHRISTIAN LE MERDY
boundedness of an appropriate functional calculus. In Section 2 below, we provide some background on Paulsen’s Theorem for the convenience of the reader. 2. Background on complete boundedness and Paulsen’s Theorem. We only give a brief account on complete boundedness and its connections with similarity problems. More information and details, as well as important developments and applications can be found in [10]. Given a Hilbert space H, we let B(H) denote the C ∗ -algebra of all bounded linear operators on H. If C is a C ∗ -algebra and n ≥ 1 is an integer, we let Mn (C) denote the C ∗ -algebra of all n × n matrices with entries in C. Let us describe the resulting norm in two important special cases. Assume first that C = B(H). Then the C ∗ -norm on Mn (B(H)) is obtained by regarding elements of Mn (B(H)) as operators on the Hilbertian direct sum H ⊕ · · · ⊕ H of n copies of H. Thus for any [Tjk ] ∈ Mn (B(H)), we have X n X n n
2 21 X
2 (2.1) [Tjk ] = sup Tjk ξk : ξk ∈ H, kξk k ≤ 1 .
j=1
k=1
k=1
Now consider the case when C = Cb (Ω) is the space of all bounded and continuous functions g : Ω → C on some topological space Ω, equipped with its sup norm. Then the C ∗ -norm on Mn (Cb (Ω)) is obtained by identifying Mn (Cb (Ω)) with the space Cb (Ω; Mn ) of bounded and continuous functions from Ω into Mn . Thus for any [gjk ] ∈ Mn (Cb (Ω)), we have n o
(2.2) [gjk ] = sup [gjk (λ)] : λ∈Ω . Mn
∗
Let H be a Hilbert space, let C be a C -algebra and let E ⊂ C be a (not necessarily closed) subspace of C. Then the space Mn (E) of n × n matrices with entries in E may be obviously regarded as embedded in Mn (C). By definition, a linear mapping u : E → B(H) is completely bounded if there exists a constant K ≥ 0 such that
[u(ajk )] ≤ K [ajk ] Mn (B(H))
Mn (E)
for any n ≥ 1 and any [ajk ] ∈ Mn (E). In that case, the least possible K is denoted by kukcb and is called the completely bounded norm of u. If the latter is ≤ 1, then we say that u is completely contractive. Obviously any completely bounded mapping u is bounded, with kuk ≤ kukcb .
SIMILARITIES OF
ω -ACCRETIVE
OPERATORS
87
Paulsen’s Theorem asserts that any completely bounded homomorphism on an operator algebra (= subalgebra of a C ∗ -algebra) is similar to a completely contractive one. More precisely, we have the following statement (see [9]), that we will use in the situation when C = Cb (Ω) for some Ω. Theorem 2.1. (Paulsen) Let H be a Hilbert space, let C be a C ∗ algebra, let A ⊂ C be a subalgebra, and consider a linear homomorphism u : A → B(H). If u is completely bounded, then there exists a bounded invertible operator S : H → H such that the linear homomorphism uS : A → B(H) defined by letting uS (a) = S −1 u(a)S for any a ∈ A is completely contractive. In particular, kS −1 u(a)Sk ≤ kak for any a ∈ A. We finally recall for further use that for any [αjk ] ∈ Mn and for any vectors ξ1 , . . . , ξn and η1 , . . . , ηn in a Hilbert space H, we have n X 21 X 12 n n
X 2 2
kξk k αjk hξk , ηj i ≤ [αjk ] Mn (2.3) kηj k . j,k=1
k=1
j=1
3. Proof of Theorem 1.1. We first introduce some notation concerning H ∞ functional calculus associated to sectorial operators (in the sense of [8], [3]). For any θ ∈ (0, π), we recall that Σθ = {z ∈ C : |Arg(z)| < θ} and we let Γθ be the counterclockwise oriented boundary of Σθ . Then we let H0∞ (Σθ ) be the space of all bounded analytic functions f : Σθ → C for which there exist two positive numbers c > 0, s > 0, such that |z|s , z ∈ Σθ . |f (z)| ≤ c 1 + |z|2s We recall that if A is a sectorial operator of type ω ∈ (0, π) and if f ∈ H0∞ (Σθ ) for some θ ∈ (ω, π), then we may define f (A) ∈ B(H) by letting Z 1 f (A) = f (z)(z − A)−1 dz, 2πi Γγ where γ ∈ (ω, θ) is an intermediate angle. (The definition of f (A) does not depend on γ by Cauchy’s Theorem.) We let A be an injective sectorial operator satisfying (1.2) for some ω ∈ (0, π2 ) and aim at proving that A is similar to an ω-accretive
88
CHRISTIAN LE MERDY
operator. Recall from [11, Theorem 2] that A is necessarily sectorial π of type ω. Changing A into A 2ω , we may assume that ω = π2 . We fix some θ ∈ ( π2 , π) and we let A0 = H0∞ (Σθ ) that we regard (by taking restrictions) as a subalgebra of Cb (Σ π2 ). Then we let A ⊂ Cb (Σ π2 ) be the 1 subalgebra linearly spanned by A0 , the function f0 (z) = 1+z , and the constant function 1. We clearly define a homomorphism u : A → B(H) by letting u(f ) = f (A) for f ∈ A0 , u(f0 ) = (1 + A)−1 , u(1) = 1, and then extending linearly. We will prove that (3.1)
u : A −→ B(H)
is completely bounded.
Taking this for granted, the conclusion goes as follows. By Paulsen’s Theorem, there exists an invertible S ∈ B(H) such that kS −1 u(f )Sk ≤ kf kCb (Σ π ) for all f ∈ A. Moreover the function f (z) = 1−z belongs to 1+z 2 −1 A and u(f ) = (1 − A)(1 + A) . Since we have 1 − z kf kCb (Σ π ) = sup : Re(z) > 0 = 1, 2 1+z we conclude that S −1 (1−A)(1+A)−1 S = (1−S −1 AS)(1+S −1 AS)−1
is a contraction.
This shows that S −1 AS is maximal accretive. To prove (3.1), we will change our sectorial functional calculus into a band sectorial functional calculus by means of the Log function. For any γ > 0, let Pγ = {λ ∈ C : |Im(λ)| < γ} and let ∆γ denote its counterclockwise oriented boundary. Let iB be the generator of the c0 -group (Ait )t , so that B should be thought as π being Log(A). Our assumption that kAit k ≤ e 2 |t| for any t ∈ R means that iB − π2 and −iB − π2 both generate contractive semigroups on H. Hence π2 − iB and π2 + iB are both maximal accretive, whence
π
π Re ( − iB)ξ, ξ ≥ 0 and Re ( + iB)ξ, ξ ≥ 0, ξ ∈ D(B). 2 2 In turn this is equivalent to say that the numerical range of B lies into the closure of P π2 , that is, (3.2)
hBξ, ξi ∈ P π2 ,
ξ ∈ D(B), kξk ≤ 1.
Let H0∞ (Pθ ) be the space of all bounded analytic functions g : Pθ → C c for which there exist a constant c > 0 such that |g(λ)| ≤ 1+|λ| 2 for any π λ ∈ Pθ . Then let γ ∈ ( 2 , θ) be an arbitrary number. Since iB − π2 and −iB − π2 both generate contractive semigroups, the function λ 7→
ω -ACCRETIVE
SIMILARITIES OF
OPERATORS
89
(λ−B)−1 is well-defined and bounded on ∆γ hence for any g ∈ H0∞ (Pθ ) we may define g(B) ∈ B(H) by letting 1 g(B) = 2πi
Z
g(λ)(λ − B)−1 dλ.
∆γ
It is easy to check (using Cauchy’s Theorem) that this definition does not depend on the choice of γ and that the mapping v : g 7→ g(B) is a linear homomorphism from H0∞ (Pθ ) into B(H). Moreover the sectorial and band functional calculi are compatible in the sense that for any f ∈ H0∞ (Σθ ), the function λ 7→ f (eλ ) belongs to H0∞ (Pθ ) and g(B) = f (A) if g(λ) = f (eλ ).
(3.3)
We refer the reader to [2] for various relationships between sectorial and band functional calculi, from which a proof of (3.3) can be extracted. However we give a direct argument for the sake of completeness. Let ϕ be the function defined by ϕ(z) = z(1+z)−2 , so that ϕ(A) = A(1+A)−2 . It is well-known that ϕ(A) has a dense range, so that we only need to prove that g(B)ϕ(A) = f (A)ϕ(A). We fix two parameters π2 < γ2 < γ1 < θ. Let λ be a complex number with Im(λ) = γ1 . Applying the Laplace formula to the semigroup (A−it )t≥0 , we have (in the strong sense) Z ∞ −1 −1 (λ − B) = i(iλ − iB) = −i eiλt A−it dt. 0
Hence using Fubini’s Theorem, we obtain ∞
−1 (λ − B) ϕ(A) = 2π
Z
1 = 2πi
Z
−1
iλt
Z
z −it ϕ(z)(z − A)−1 dz dt
e 0
Γγ2 Z ∞
−i
Γγ2
iλt −it
e z
dt ϕ(z)(z − A)−1 dz
0
whence 1 (λ − B) ϕ(A) = 2πi −1
Z Γγ2
1 ϕ(z)(z − A)−1 dz. λ − Log(z)
The latter idendity can be proved as well if Im(λ) = −γ1 hence holds true for any λ ∈ ∆γ1 . Using Fubini’s Theorem again and Cauchy’s
90
CHRISTIAN LE MERDY
Theorem, we therefore deduce that Z 1 g(B)ϕ(A) = g(λ)(λ − B)−1 ϕ(A) dλ 2πi ∆γ1 Z 1 2 Z 1 g(λ) ϕ(z)(z − A)−1 dz dλ = 2πi λ − Log(z) ∆ γ1 Γγ2 1 2 Z Z 1 = g(λ) dλ ϕ(z)(z − A)−1 dz 2πi λ − Log(z) Γγ2 ∆γ1 Z 1 = g(Log(z))ϕ(z)(z − A)−1 dz 2πi Γγ2 = f (A)ϕ(A), which concludes the proof of (3.3). We let B be equal to H0∞ (Pθ ) regarded as a subalgebra of Cb (P π2 ). To prove (3.1), it will suffice to show that (3.4)
v : B −→ B(H)
is completely bounded.
Indeed since the exponential function is a holomorphic bijection from P π2 onto Σ π2 , it follows from (3.3) and the definition of the matrix norms on A and B (see (2.2)) that if v is completely bounded, then u|A0 is completely bounded as well, with ku|A0 kcb ≤ kvkcb . However A0 has codimension 2 in A hence the complete boundednes of u|A0 implies that of u on A. We now come to the heart of the proof, which consists in showing that for an operator B whose spectrum is included in P π2 , the condition (3.2) implies (3.4). That (3.2) implies the boundedness of v is a recent result of Crouzeix and Delyon ([5]) and our proof of the complete boundedness of v will essentially be a repetition of their arguments, up to some adequate matrix norm manipulations. Before embarking into computations, we notice that (3.2) is equivalent to the following real/imaginary parts decomposition for B: π (3.5) B = C + iD, with C = C ∗ , D = D∗ , kDk ≤ . 2 In this decomposition, C is a possibly unbounded self-adjoint operator with D(C) = D(B). Let (E(s))s be the resolution of the identity for C and for any integer m ≥ 1, let Z Cm = s dE(s) and Bm = Cm + iD. (−m,m)
SIMILARITIES OF
ω -ACCRETIVE
OPERATORS
91
Then Cm is a bounded self-adjoint operator hence Bm is a bounded operator whose numerical range lies in P π2 . Moreover for any λ ∈ / P π2 , we have (3.6)
(λ − Bm )−1 −→ (λ − B)−1
strongly.
Indeed, (λ − Bm )−1 − (λ − B)−1 = (λ − Bm )−1 (Cm − C)(λ − B)−1 , we have Cm ξ → Cξ for any ξ ∈ D(B) = D(C), and since the operators π ± iBm are maximal accretive, we have a uniform estimate 2 (3.7)
k(λ − Bm )−1 k ≤ d(λ, P π2 ),
m ≥ 1.
Next, by Lebesgue’s Theorem, it follows from (3.6) and (3.7) that g(Bm ) → g(B) strongly for any g ∈ B. Thus for any n ≥ 1 and any [gjk ] ∈ Mn (B), we have
[gjk (B)] ≤ lim sup [gjk (Bm )] m
To prove the complete boundedness of v, it therefore suffices to prove that the mappings g 7→ g(Bm ) are uniformly completely bounded. To achieve this goal we shall now assume that B is bounded and shall prove that 2 (3.8) kvkcb ≤ √ + 2. 3 Let γ ∈ ( π2 , θ) be an arbitrary intermediate angle. Then according to [5, (5)] (and its proof), we may write v(g) = g(B) = v1γ (g) + v2γ (g) for any g ∈ B = H0∞ (Pθ ), with Z +∞ 1 γ v1 (g) = g(x) (x + 2γi − B ∗ )−1 − (x − 2γi − B ∗ )−1 dx; 2πi −∞ Z 1 γ v2 (g) = g(λ) (λ − B)−1 − (λ − B ∗ )−1 dλ. 2πi ∆γ Moreover it is easy to check that for any x ∈ R, one has (x + 2γi − B ∗ )−1 − (x − 2γi − B ∗ )−1 = −4γi (x + 2γi − B ∗ )−1 (x − 2γi − B ∗ )−1 −1 = −4γi M (x) − iN (x) , where M (x) and N (x) are self-adjoint operators defined by M (x) = (x − C)2 − D2 + 4γ 2
and
N (x) = CD + DC − 2xD.
92
CHRISTIAN LE MERDY
(The boundedness of C allows this real/imaginary parts decomposition.) It follows from (3.5) that π 2 (3.9) M (x) ≥ (x − C)2 + 3 . 2 1
1
In particular, M (x) is invertible and with Q(x) = M (x)− 2 N (x)M (x)− 2 , we may write −1 1 1 (x+2γi−B ∗ )−1 −(x−2γi−B ∗ )−1 = −4γi M (x)− 2 1−iQ(x) M (x)− 2 . Let n ≥ 1 be an integer and let [gjk ] be an element of Mn (B) with norm ≤ 1. According to (2.2), this simply means that
[gjk (λ)] ≤ 1, (3.10) λ ∈ Pπ . Mn
2
We let ξ1 , . . . , ξn , η1 , . . . , ηn be arbitrary elements of H. Then n X
v1γ (gjk )ξk , ηj
j,k=1
Z n X
−1 1 1 −2γ +∞ gjk (x) M (x)− 2 1 − iQ(x) M (x)− 2 ξk , ηj dx = π −∞ j,k=1 Z n −2γ +∞ X
−1 1 1 gjk (x) 1 − iQ(x) M (x)− 2 ξk , M (x)− 2 ηj dx . = π −∞ j,k=1 Applying (2.3) and (3.10), we obtain that X n γ v1 (gjk )ξk , ηj j,k=1
2γ ≤ π
Z
+∞X
−∞
1 X
1
M (x)− 12 ηj 2 2 dx.
1 − iQ(x) −1 M (x)− 21 ξk 2 2
k
j
−1 is a contraction Since Q(x) is self-adjoint, the operator 1 − iQ(x) for any x ∈ R hence applying Cauchy-Schwarz, we finally obtain that X n γ v1 (gjk )ξk , ηj j,k=1
21 X Z +∞ 21 Z
1 1 2γ X +∞ 2 2
M (x)− 2 ξk dx
M (x)− 2 ηj dx . ≤ π −∞ −∞ j k
SIMILARITIES OF
ω -ACCRETIVE
OPERATORS
93
Now observe that for any ξ ∈ H, we have Z
+∞
M (x)− 12 ξ 2 dx =
−∞
≤
Z
+∞
hM (x)−1 ξ, ξi dx
−∞ Z +∞ D
(x − C)2 + 3
−∞
π 2 −1 E ξ, ξ dx 2
by (3.9). Moreover using the spectral representation of C we see that the latter integral is equal to Z
+∞
−∞
kξk2
2 dx = √ kξk2 . 2 3 x2 + 3 π2
Combining with the previous estimate, this yields n 1 X 1 X γ 4γ X 2 2 2 2 √ kξ k kη k . v (g )ξ , η ≤ k j 1 jk k j π 3 j k j,k=1 In view of the definition of matrix norms on B(H) (see (2.1)), we deduce (3.11)
γ
[v1 (gjk )] ≤ 4γ √ . π 3
We now turn to an estimate for v2γ . We rewrite the definition of the latter mapping as Z γ v2 (g) = g(λ)T (λ) |dλ|, ∆γ
1 where T (λ) is equal to 2πi (λ − B)−1 − (λ − B ∗ )−1 if Im(λ) = −γ and is equal to its opposite if Im(λ) = γ. The key point is that T (λ) is a nonnegative operator for any λ ∈ ∆γ . Indeed assume for example that Im(λ) = −γ. Then 1 (λ − B)−1 − (λ−B ∗ )−1 2πi 1 = (λ − B)−1 (2iγ + B − B ∗ )(λ − B ∗ )−1 2πi 1 = (λ − B)−1 (γ + D)(λ − B ∗ )−1 , π
94
CHRISTIAN LE MERDY
which is nonnegative by (3.5). Then arguing as above, we obtain that for any vectors ξ1 , . . . , ξn , and η1 , . . . , ηn ∈ H, we have X 12 Z n n γ
o X
2
v2 (gjk )ξk , ηj ≤ sup [gjk (λ)] T (λ)ξk |dλ| λ∈Pγ
j,k=1
k
×
X Z
∆γ
21
2
T (λ)ηj |dλ| .
∆γ
j
Now observe that since B is bounded, the function λ 7→ (λ−B)−1 −(λ− R 1 (λ−B)−1 −(λ−B ∗ )−1 dλ = 2 B ∗ )−1 is integrable on ∆γ and that 2πi ∆γ by Cauchy’s Theorem. Hence for any ξ ∈ H, we have Z Z
T (λ)ξ 2 |dλ| = 1 (λ−B)−1 −(λ−B ∗ )−1 ξ, ξ dλ = 2kξk2 . 2πi ∆γ ∆γ Combining with the above estimate, we obtain that n
γ
o
[v2 (gjk )] ≤ 2 sup [gjk (λ)] . λ∈Pγ
Since
lim
γ→ π2
n n
o
o
sup [gjk (λ)] = sup [gjk (λ)] ≤ 1, λ∈Pγ
λ∈P π 2
we finally deduce that n
o
[v(gjk )] ≤ inf [v1γ (gjk )] + [v2γ (gjk )] ≤ √2 + 2, γ> π2 3 which concludes our proof of (3.8). Remark 3.1. Two results analogous to the one in [5] appear in [6] and [4]. On the one hand, it is shown in [6] that if Ω ⊂ C is bounded and convex and if B is a bounded operator on H whose numerical range lies in Ω, then the analytic functional calculus associated to B is bounded with respect to the norm induced by Cb (Ω). On the other hand, it is shown in [4] that if A is an ω-accretive operator on H, then its analytic functional calculus is bounded with respect to the norm induced by Cb (Σω ). In the two cases, it it actually possible to show that these bounded functional calculi are completely bounded. If we apply Paulsen’s Theorem to the functional calculus considered in [4] (sectorial case), we recover Corollary 1.2. Acknowledgements. This research was carried out while I was visiting the Centre for Mathematics and its Applications at the Australian National University in Canberra. It is a pleasure to thank the CMA for
SIMILARITIES OF
ω -ACCRETIVE
OPERATORS
95
its warm hospitality. I am also grateful to Alan McIntosh for pointing out to me the paper [5]. References [1] D. Albrecht, X. Duong, and A. McIntosh, Operator theory and harmonic analysis, Proc. of CMA, Canberra 34 (1996), 77-136. [2] K. Boyadzhiev, and R. deLaubenfels, Spectral theorem for unbounded strongly continuous groups on a Hilbert space, Proc. Amer. Math. Soc. 120 (1994), 127-136. [3] M. Cowling, I. Doust, A. McIntosh, and A. Yagi, Banach space operators with a bounded H ∞ functional calculus, J. Austr. Math. Soc. (Series A) 60 (1996), 51-89. [4] M. Crouzeix, Une majoration du type von Neumann pour les fractions rationnelles d’op´erateurs sectoriels, C. R. Acad. Sci. Paris 330, S´erie I (2000), 513-516. [5] M. Crouzeix, B. Delyon, Some estimates for analytic functions of band or sectorial operators, Preprint (2001). [6] B. Delyon, and F. Delyon, Generalization of von Neumann’s spectral sets and integral representations of operators, Bull. Soc. Math. France 127 (1999), 25-42. [7] C. Le Merdy, The similarity problem for bounded analytic semigroups on Hilbert space, Semigroup Forum 56 (1998), 205-224. [8] A. McIntosh, Operators which have an H ∞ functional calculus, Proc. of CMA, Canberra 14 (1986), 210-231. [9] V. Paulsen, Completely bounded homomorphisms of operator algebras, Proc. Amer. Math. Soc. 02 (1984), 225-228. [10] G. Pisier, Similarity problems and completely bounded maps, Lecture Notes 1618, Springer Verlag, 1996. [11] J. Pr¨ uss, and H. Sohr, On operators with bounded imaginary powers in Banach spaces, Math. Z. 203 (1990), 429-452. [12] A. Simard, Counterexamples concerning powers of sectorial operators on a Hilbert space, Bull. Austr. Math. Soc. 60 (1999), 459-458. ´partement de Mathe ´matiques, Universite ´ de Franche-Comte ´, 25030 De Besanc ¸ on Cedex, France E-mail address: [email protected]
JACOBIANS ON LIPSCHITZ DOMAINS OF R2 ZENGJIAN LOU Dedicated to Professor Alan McIntosh on the occasion of his 60th birthday Abstract. In this note we prove estimates of Jacobian determinants of Du on strongly Lipschitz domains Ω in R2 . The theorem consists of two parts: one is an estimate in terms of the BM Or (Ω) norm for u in the Sobolev space W 1,2 (Ω, R2 ) with boundary zero, and another is an estimate in terms of the BM Oz (Ω) norm for u in W 1,2 (Ω, R2 ) with no boundary conditions.
1. Introduction Jacobian determinant estimates were first studied by M¨ uller in [Mu]. In [CLMS], Coifman, Lions, Meyer and Semmes’ Theorems II.1 and R III.2 imply that for b ∈ L2loc (R2 ), supu R2 b det Du dx is equiva∂u lent to the BM O(R2 ) norm of b, where det Du(x) = ∂xkj , the supremum is taken over all u in the Sobolev space W 1,2 (R2 , R2 ) with kDui kL2 (R2 ,R2 ) ≤ 1. The aim of this note is to consider an extension of this result to domains in R2 . As a main result (Theorem 2.1), we give R estimates of supu Ω b det Du dx when Ω is a strongly Lipschitz domain of R2 , where the superemum is taken over all u in the Sobolev space W 1,2 (Ω, R2 ) or W01,2 (Ω, R2 ) (the closure of C0∞ (Ω, R2 ) in W 1,2 (Ω, R2 )) with kDui kL2 (Ω,R2 ) ≤ 1. In the sequel, Ω will denote a strongly Lipschitz domain - an assumption which is enough to ensure (1) the existence of a bounded extension map from W 1,2 (Ω, R2 ) to W 1,2 (R2 , R2 ), and (2) the existence of a bounded extension map from BM Or (Ω) to BM O(R2 ), where BM Or (Ω) is the space of locally integrable functions with Z 1 kf kBM Or (Ω) = sup |f (x) − fQ | dx < ∞, |Q| Q Q⊂Ω 1991 Mathematics Subject Classification. Primary 42B30. Secondary 35J45. Key words and phrases. Jacobian determinant, Lipschitz domain, BM Oz (Ω), BM Or (Ω), coercivity. 96
JACOBIANS ON LIPSCHITZ DOMAINS OF R2
97
R 1 here fQ = |Q| f (x) dx, the supremum is taken over all cubes Q in Q the domain Ω. In [CKS], two Hardy spaces are defined on bounded domains Ω, one which is reasonably speaking the largest, and the other which in a sense is the smallest. The largest, Hr1 (Ω), arises by restricting to Ω arbitrary elements of H1 (R2 ). The other, Hz1 (Ω), arises by restricting ¯ Norms on these to Ω elements of H1 (R2 ) which are zero outside Ω. spaces are defined as following kf kHr1 (Ω) = inf kF kH1 (R2 ) , the infimum being taken over all the functions F ∈ H1 (R2 ) such that F |Ω = f , kf kHz1 (Ω) = kF kH1 (R2 ) , where F is the zero extension of f to R2 . From [C], the dual of Hz1 (Ω) is BM Or (Ω) and the dual of Hr1 (Ω) is BM Oz (Ω), where BM Oz (Ω) is the space of all functions in BM O(R2 ) ¯ equipped with the norm kf kBM Oz (Ω) = kf kBM O(R2 ) . supported in Ω, 2. The Main Theorem and Its Proof In [CLMS, Theorems II.1 and III.2], among other results, Coifman, Lions, Meyer and Semmes established the following: (A) If u ∈ W 1,2 (R2 , R2 ) then det Du belongs to the Hardy space 1 H (R2 ) and kdet Duk
H1 (R2 )
≤C
2 Y
kDui kL2 (R2 ,R2 )
(2.1)
b E · F dx,
(2.2)
i=1
for some absolute constants C. (B) b ∈ L2loc (R2 ) Z kbkBM O(R2 ) ∼ sup E,F
R2
where the supremum is taken over all E, F ∈ L2 (R2 , R2 ) with div E = curl F = 0 and kEkL2 (R2 ,R2 ) , kF kL2 (R2 ,R2 ) ≤ 1. We will see that (A) and (B) yield the following equivalence Z kbkBM O(R2 ) ∼ sup b det Du dx, (2.3) u
R2
where the supremum is taken over all u = (u1 , u2 ) ∈ W 1,2 (R2 , R2 ) with kDui kL2 (R2 ,R2 ) ≤ 1, i = 1, 2.
98
ZENGJIAN LOU
Suppose that E and F satisfy the conditions of (B). From Theorems 2.9 and 3.1 in [GR], there exist ϕ, ψ ∈ W 1,2 (R2 ) such that ∂ϕ ∂ϕ E = curl ϕ = , − ∂x2 ∂x1 and ∂ψ ∂ψ F = Dψ = , . ∂x1 ∂x2 Define u = (ϕ, ψ). Then u ∈ W 1,2 (R2 , R2 ) with kDui kL2 (R2 ,R2 ) ≤ 1, i = 1, 2, and det Du = −E · F. Thus (2.2) implies that Z kbkBM O(R2 ) ≤ C sup b det Du dx . u
R2
Conversely, applying (2.1) and the duality H1 (R2 )∗ = BM O(R2 ), we have R b det Du dx ≤ CkbkBM O(R2 ) kdet DukH1 (R2 ) R2 ≤ CkbkBM O(R2 )
2 Y
kDui kL2 (R2 ,R2 )
i=1
≤ CkbkBM O(R2 ) if kDui kL2 (R2 ,R2 ) ≤ 1. A natural question to ask is: under what conditions does (2.3) hold on domains Ω of R2 ? As a main theorem of this note, we solve this problem for strongly Lipschitz domains in R2 . Theorem 2.1. Let Ω be a strongly Lipschitz domain in R2 . (1) If b ∈ BM Oz (Ω), then we have equvalence Z kbkBM Oz (Ω) ∼ sup b det Du dx, u
(2.4)
Ω
the supremum being taken over all u = (u1 , u2 ) ∈ W 1,2 (Ω, R2 ) with kDui kL2 (Ω,R2 ) ≤ 1, i = 1, 2. (2) If b ∈ BM Or (Ω), then Z kbkBM Or (Ω) ∼ sup b det Du dx, (2.5) u
Ω
the supremum being taken over all u = (u1 , u2 ) ∈ W01,2 (Ω, R2 ) with kDui kL2 (Ω,R2 ) ≤ 1, i = 1, 2. The implicit constants in (2.4) and (2.5) depend only on the domain Ω.
JACOBIANS ON LIPSCHITZ DOMAINS OF R2
99
To prove Theorem 2.1, we need the following Lemmas 2.2 - 2.4. The proof of Lemma 2.2 is given at the end of this section. Lemma 2.3 is the two-dimensional case of Theorem 3.1 in Section 3. Lemma 2.4 is a special case of an extension theorem by Jones in [J, Theorem 1]. We also need the the following seminorm defined in [Z] Z 1 kbkBM OH (Ω) = sup |b − bQ | dx , |Q| Q Q where the supremum is taken over all cubes Q with 2Q ⊂ Ω. Lemma 2.2. Let Ω be an open domain in R2 . For b ∈ L2loc (Ω) Z kbkBM OH (Ω) ≤ C sup b det Du dx u
Ω
for a constant C independent of b, the supremum being taken over all u = (u1 , u2 ) ∈ W01,2 (Ω, R2 ) with kDui kL2 (Ω,R2 ) ≤ 1, i = 1, 2. Lemma 2.3. Let Ω ⊂ R2 be a strongly Lipschitz domain and let b be a locally integrable function on Ω. Then kbkBM Or (Ω) ∼ kbkBM OH (Ω) , where the implicit constants are independent of b. Lemma 2.4. Let Ω be a strongly Lipschitz domain in R2 and let b ∈ BM Or (Ω). Then there exists B ∈ BM O(R2 ) such that B|Ω = b and kBkBM O(R2 ) ≤ CkbkBM Or (Ω) for some constants C independent of B and b. Proof of Theorem 2.1. (1) Suppose b ∈ BM Oz (Ω). Define ( b in Ω; B= 0 in R2 \ Ω. By the definition of BM Oz (Ω), B ∈ BM O(R2 ) and kBkBM O(R2 ) = kbkBM Oz (Ω) .
(2.6)
Since Ω is a bounded strongly Lipschitz domain, u ∈ W 1,2 (Ω, R2 ) can be extended to U ∈ W 1,2 (R2 , R2 ) with kDUi kL2 (R2 ,R2 ) ≤ CkDui kL2 (Ω,R2 ) ,
(2.7)
100
ZENGJIAN LOU
where the constant C depends only on the Lipschitz constant of Ω (see, for example, Proposition 4.12 in [HLMZ]). Therefore (2.6), (2.7) and (2.1) give Z
Z b det Du dx =
B det DU dx R2
Ω
≤ kBkBM O(R2 ) kdet DU kH1 (R2 ) ≤ CkbkBM Oz (Ω)
2 Y
kDUi kL2 (R2 ,R2 )
i=1
≤ CkbkBM Oz (Ω)
2 Y
kDui kL2 (Ω,R2 )
i=1
≤ CkbkBM Oz (Ω) if kDui kL2 (Ω,R2 ) ≤ 1, where C depends only on the domain Ω. We now prove the converse. Let b ∈ BM Oz (Ω) and define B as above. Suppose U ∈ W 1,2 (R2 , R2 ) and kDUi kL2 (R2 ,R2 ) ≤ 1. Let u = U |Ω , then kDui kL2 (Ω,R2 ) ≤ 1. So (2.3) and (2.6) yield kbkBM Oz (Ω) = kBkBM O(R2 ) Z ≤C
B det DU dx
sup U ∈W 1,2 (R2 ,R2 ),kDUi kL2 ≤1
R2
Z =C
b det Du dx
sup u=U |Ω ,U ∈W 1,2 (R2 ,R2 ),kDUi kL2 ≤1
Ω
Z ≤C
sup u∈W 1,2 (Ω,R2 ),kDui kL2 ≤1
b det Du dx. Ω
(2) Let B ∈ BM O(R2 ) is an extension of b ∈ BM Or (Ω). For u ∈ W01,2 (Ω, R2 ), define ( u in Ω; U= 0 in R2 \ Ω. Then U ∈ W 1,2 (R2 , R2 ) with kU kW 1,2 (R2 ,R2 ) = kukW 1,2 (Ω,R2 ) .
JACOBIANS ON LIPSCHITZ DOMAINS OF R2
101
By duality H1 (R2 )∗ = BM O(R2 ), (2.1) and Lemma 2.4, we have Z Z b det Du dx = B det DU dx Ω
R2
= kBkBM O(R2 ) kdet DU kH1 (R2 ) ≤ CkbkBM Or (Ω)
2 Y
kDUi kL2 (R2 ,R2 )
i=1
= CkbkBM Or (Ω)
2 Y
kDui kL2 (Ω,R2 )
i=1
≤ CkbkBM Or (Ω) for all u ∈ W01,2 (Ω, R2 ) with kDui kL2 (Ω,R2 ) ≤ 1, i = 1, 2, where the constant C depends only on the domain Ω. The proof of the reversed inequality in (2.5) follows from Lemmas 2.2 and 2.3. Theorem 2.1 is proved. We now prove Lemma 2.2, its proof uses the following result of Ne˘cas [N, Lemma 7.1, Chapter 3]. Lemma 2.5. Let Ω be a bounded strongly Lipschitz domain in RN . Then the divergence operator isR a (continuous) map from W01,2 (Ω, RN ) onto L20 (Ω) = {f ∈ L2 (Ω) : Ω f dx = 0}. That is, there exists a constant C depending only on the domain Ω and the dimension N such that for any f ∈ L20 (Ω), there exists ϕ ∈ W01,2 (Ω, RN ) such that f = div ϕ and kDϕkL2 (Ω,RN ) ≤ Ckf kL2 (Ω) . Proof of Lemma 2.2. Suppose b ∈ L2loc (Ω). We will show that there exists u = (u1 , u2 ) ∈ W01,2 (Ω, R2 ) with kDui kL2 (Ω,R2 ) ≤ 1 for i = 1, 2, and supp (det Du) ⊂ Q such that for all cubes Q with 2Q ⊂ Ω, 1/2 Z Z 1 2 ≤ C b det Du dx , |b − bQ | dx (2.8) |Q| Q Q R 1 where bQ = |Q| b dx, C is a constant independent of Q, b and u. Q R Let h = b − bQ , then h ∈ L2 (Q) with Q h dx = 0. Using Lemma 2.5 with Ω = Q, there exists ϕ = (ϕ1 , ϕ2 ) ∈ W01,2 (Q, R2 ) and an absolute constant C0 such that h = div ϕ
102
ZENGJIAN LOU
and kDϕkL2 (Q,R2 ) ≤ C0 khkL2 (Q) .
(2.9)
So khk2L2 (Q)
Z =
h div ϕ dx Q
Z
Z ∂ϕ1 ∂ϕ2 = h dx + h dx Q ∂x1 Q ∂x2 Z ∂ϕ i ≤ 2 max h dx 1≤i≤2 Q ∂xi Z ∂ϕ i0 dx = 2 h Q ∂xi0
(2.10)
for some choice of i0 (i0 = 1 or 2). Assuming without loss of generality that i0 = 1 in (2.10). To prove (2.8), we need only to show that there exists u ∈ W01,2 (Ω, R2 ) with conditions stated above and a constant C (independent of Q, ϕ and u) such that Z Z ∂ϕ1 −1 1/2 dx ≤ C|Q| h det Du dx . (2.11) hkhkL2 (Q) ∂x1 Q Q Set u1 =
ϕ1 . C0 khkL2 (Q)
It is obvious that u1 ∈ W01,2 (Q) and kDu1 kL2 (Q,R2 )
≤ 1 by (2.9). Let ψ0 ∈ C0∞ (R2 ) such that ( 1 on [−1, 1]2 ; ψ0 = 0 outside [−2, 2]2 . Define u2 = γC0 |Q|−1/2 (x2 − x02 )ψQ (x), x−x0 where ψQ (x) = ψ0 l(Q)/2 , x0 = (x01 , x02 ) denotes the center of the cube Q, γ > 0 is a normalization constant (independent of x0 and l(Q)) so that kDu2 kL2 (R2 ,R2 ) ≤ 1. It is obvious that u2 ∈ C0∞ (2Q). Let u = (u1 , u2 ). By a simple computation, we get det Du = γ|Q|−1/2 khk−1 L2 (Q) So (2.11) follows. Lemma 2.2 is proved.
∂ϕ1 ∂x1
in Q.
JACOBIANS ON LIPSCHITZ DOMAINS OF R2
103
3. The Equivalence of Two BM O Seminorms In [Z] Zhang asked if the two seminorms kbkBM Or (Ω) and kbkBM OH (Ω) are equivalent under suitable conditions on domains Ω in RN . The following theorem gives a positive answer. No smoothness conditions are needed on the domains. In addition, we will see that the equivalence of kbkBM Or (Ω) and kbkBM OH (Ω) implies that Hz1 (Ω) can be decomposed into a sum of atoms with supports away from boundaries of the domains (Proposition 3.2). Theorem 3.1. Let Ω ⊂ RN ( N ≥ 2) be a strongly Lipschitz domain and let b be a locally integrable function on Ω. Then kbkBM Or (Ω) ∼ kbkBM OH (Ω) , where the implicit constants are independent of the function b. Proof. It is obvious that kbkBM OH (Ω) ≤ kbkBM Or (Ω) . Now we prove kbkBM Or (Ω) ≤ CkbkBM OH (Ω) . We will only give the proof for the case of RN for N = 2 using ideas of Jones [J]. The case for RN (N 6= 2) is similar. In fact, from Jones’ extension theorem [J, Theorem 1] we need only to prove that there exists a constant C independent of Q and b such that for all cubes Q ⊂ Ω, Z 1 |b − bQ | dx ≤ CkbkBM OH (Ω) . (3.1) |Q| Q LetSE = {Qk } be the dyadic Whitney decomposition of Q, then Q = k Qk and Qj ∩ Qk = ∅, j 6= k; (3.2) 1≤
d(Qk , Qc ) ≤ 4 · 21/2 ; l(Qk )
(3.3)
1 l(Qj ) ≤ ≤ 4 if Qj ∩ Qk 6= ∅ (3.4) 4 l(Qk ) (see, for example, Stein’s book [S1, page 167] for more information on Whitney decompositions). For m = 1, 2, ..., let d(x, Qc ) −m −m+1 Am = x ∈ Q : 2 ≤ <2 l(Q) and Fm = {Qj ∈ E : Qj ∩ Am 6= ∅}. To prove (3.1), we need the following two claims:
104
ZENGJIAN LOU
Claim (A). X
|Qj | ≤ 40 · 2−m |Q|
Qj ∈Fm
for all m = 1, 2, .... Claim (B). If b ∈ BM OH (Ω), then |bQj − bQ0 | ≤ CmkbkBM OH (Ω) for all Qj ∈ Fm , m = 1, 2, ..., where Q0 be the Whitney cube that contains the center of Q and C is a constant independent of b and Qj . We shall give the proofs of these claims later on and we first prove (3.1) admitting them. Since Qj ∈ E, j = 1, 2, ..., it is obvious that Z 1 (3.5) |b − bQj | dx ≤ kbkBM OH (Ω) . |Qj | Qj By (3.5), Claims (A) and (B), we have 1 |Q|
Z |b − bQ0 | dx ≤ Q
∞ X X m=1 Qj ∈Fm
∞ X X |Qj | = |Q| m=1 Q ∈F j
m
1 |Q|
Z
|bQj
|b − bQ0 | dx Qj
1 − bQ0 | + |Qj |
!
Z |b − bQj | dx Qj
∞ X X |Qj | ≤ (CmkbkBM OH (Ω) + kbkBM OH (Ω) ) |Q| m=1 Q ∈F j
≤
∞ X
m
40 · 2−m (Cm + 1)kbkBM OH (Ω)
m=1
≤CkbkBM OH (Ω) . Therefore Z Z 1 1 |b − bQ | dx ≤ (|b − bQ0 | + |bQ0 − bQ |) dx |Q| Q |Q| Q Z 1 ≤ |bQ0 − bQ | + |b − bQ0 | dx |Q| Q Z 2 ≤ |b − bQ0 | dx |Q| Q ≤ CkbkBM OH (Ω) . This gives (3.1).
JACOBIANS ON LIPSCHITZ DOMAINS OF R2
105
The proof of claims (A) was given in [J, page 47 (3.8)]. To prove Claim (B) we need the following Claim (C). If b ∈ BM OH (Ω) and Qj , Qk ∈ E have touching edges for j 6= k, then |bQj − bQk | ≤ CkbkBM OH (Ω) for an absolute constant C. Proof. Suppose that Qj and Qk touch and satisfy (3.4). Dividing (3.4) into two cases, case (a): 1 l(Qj ) ≤ ≤1 4 l(Qk )
(3.6)
l(Qj ) ≤ 4. l(Qk )
(3.7)
and case (b): 1<
For case (a), constructing cubes Rj , Rk and Rjk such that 1) Rj ⊂ Qj , Rk ⊂ Qk , l(Rj ) = 12 l(Qj ), l(Rk ) = 21 l(Qk ) and Rj , Rk touch; 2) Rj , Rk ⊂ Rjk , l(Rjk ) = l(Rj ) + l(Rk ); 3) l(Rjk ) ≤ d(Rjk , Qc ). Since Qj , Qk ∈ E touch and Qj ∩ Qk = ∅, it is easy to find cubes Rj , Rk and Rjk satisfying 1) and 2). In order for the cube Rjk to satisfy 3) we need only to choose Rjk such that d(Qj , Qc ) ≤ d(Rjk , Qc ) and d(Qk , Qc ) ≤ d(Rjk , Qc ). Therefore 1 1 l(Rjk ) = l(Qj ) + l(Qk ) 2 2 1 1 d(Qj , Qc ) + d(Qk , Qc ) ≤ 2 2 ≤ d(Rjk , Qc ). From 1) and (3.5) we have |bRj
Z 1 (b − bQj ) dx − bQj | = |Rj | Rj Z 4 ≤ |b − bQj | dx |Qj | Qj
(3.8)
≤ 4kbkBM OH (Ω) . Similarly we get |bRk − bQk | ≤ 4kbkBM OH (Ω) .
(3.9)
106
ZENGJIAN LOU
By (3.6), 1) and 2) we know that l(Qj ) l(Qk ) + 2 2 5 ≤ l(Qj ) = 5l(Rj ) 2
l(Rjk ) =
and
(3.10)
l(Qj ) l(Qk ) + (3.11) 2 2 ≤ l(Qk ) = 2l(Rk ). ⊂ Ω and l(Rjk ) ≤ d(Rjk , Ωc ). Then (3.10) yields Z 1 |bRj − bRjk | = (b − bRjk ) dx |Rj | Rj Z 25 (3.12) |b − bRjk | dx ≤ |Rjk | Rjk l(Rjk ) =
Note that Rjk
≤ 25kbkBM OH (Ω) . By (3.11), similar to (3.12) we have |bRk − bRjk | ≤ 4kbkBM OH (Ω) .
(3.13)
Combining (3.8), (3.9), (3.12) and (3.13) we get |bQj − bQk | ≤ |bQj − bRj | + |bRj − bRjk | + |bRjk − bRk | + |bRk − bQk | ≤ 37kbkBM OH (Ω) for all Qj , Qk satisfying (3.6). For case (b), repeat the process above we obtain |bQj − bQk | ≤ 37kbkBM OH (Ω) for all Qj , Qk satisfying (3.7). Therefore for all Qj , Qk ∈ E (j 6= k) have touching edges |bQj − bQk | ≤ 37kbkBM OH (Ω) . This proves Claim C.
Proof of Claim (B). The proof is similar to the argument in [J]. Let xj ∈ Qj ∈ Fm , xQ be the center of the cube Q. Then (3.3) and (3.4) show that there are at most 50 cubes Qk ∈ Fm intersect the line segment xj xQ and at most m sets Ai , i = 1, 2, ..., m, intersect xj xQ . So from Claim (C), we have |bQj − bQ0 | ≤ CmkbkBM OH (Ω) . Claim (B) is proved. The proof of Theorem 3.1 is finished completely.
JACOBIANS ON LIPSCHITZ DOMAINS OF R2
107
Remark. It should be added that at the time Theorem 3.1 was finished, the author was unfortunately unaware of a similar work in [RR] (with a different proof). Thanks go to P. Schvartsman (Department of Mathematics, Techion, 3200 Haifa, Israel) for pointing this out to him. In addition, Auscher and Russ also proved Theorem 3.1 by using duality [AR, Theorem 6]. We know that any f in Hz1 (Ω) has a decomposition (see [CKS], [C] and [JSW] for bounded domains, [AR] for unbounded domains) f=
∞ X
λ k ak
k=0 1 with k |λk | ≤ Ckf kHz1 (Ω) , where the ak ’s are H R z (Ω)-atoms: there exist cubes Qk ⊂ Ω such that supp ak ⊂ Qk , Qk ak dx = 0 and kak kL2 (Qk ) ≤ |Qk |−1/2 . In the following proposition we prove that the supports of these atoms can be away from the boundary of Ω by using 1 Theorem 3.1. Let Hz,2at (Ω) denote the space of f ∈ Hz1 (Ω) which can be decomposed into a sum of Hz1 (Ω)-atoms supported in cubes Q with 2Q ⊂ Ω.
P
Proposition 3.2. For a strongly Lipschitz domain Ω in RN 1 Hz1 (Ω) = Hz,2at (Ω). 1 (Ω) ⊂ Hz1 (Ω). So to prove the Proof. Obviously we have that Hz,2at 1 (Ω)∗ ⊂ BM Or (Ω) = proposition we only need to show that Hz,2at BM OH (Ω) := {f : kf kBM OH (Ω) < ∞}. Suppose that L is a bounded 1 (Ω). Follow the lines of the proof of Theorem linear functional on Hz,2at 1 in [S2, Chapter IV] or Theorem 2.5 in [GHL], we see that there exists a function g ∈ BM OH (Ω) such that Z L(f ) = f (x)g(x) dx Ω 1 for all f ∈ Hz,2at (Ω). The proof is finished.
4. An Application In this section we give an application of Theorem 3.1 which improves a coercivity result by Zhang. In [Z], Zhang studied the coercivity of strongly elliptic quadratic forms with measurable coefficients, defined on a bounded domain Ω in R2 with Lipschitz boundary, Z ∂ui ∂uj a(u, Ω) = Aij (x) dx, u ∈ W01,2 (Ω, R2 ), α,β ∂x ∂x α β Ω
108
ZENGJIAN LOU
∞ where Aij α,β ∈ L (Ω) and satisfy Legendre-Hadamard condition i j 2 2 Aij α,β (x)ξα ξβ η η ≥ c|ξ| |η| . i j From [M], Aij α,β Pα Pβ can be written in the form ij i j i j Aij α,β Pα Pβ = Bα,β Pα Pβ + b(x) det P,
(4.1)
ij where P ∈ M 2×2 , the set of real-valued 2 × 2 matrices, Bα,β ∈ L∞ (Ω) and satisfying ij (4.2) C1 |P |2 ≤ Bα,β (x)Pαi Pβj ≤ C2 |P |2 , for constants C1 , C2 > 0. As one of the main results in [Z], the following theorem was proved by Zhang which establishes the necessary condition such that a(u, Ω) ≥ 0.
Theorem 4.1. Suppose that Ω ⊂ R2 is a strongly Lipschitz domain, 2 Aij α,β : Ω → R is measurable for 1 ≤ i, j, α, β ≤ 2, such that (4.1) holds, ij are measurable functions satisfying (4.2) where b ∈ BM Or (Ω) and Bα,β for constants 0 < C1 < C2 . Then there exists a constant C3 depending only on C2 such that a(u, Ω) ≥ 0 for all u ∈ W01,2 (Ω, R2 ) implies that kbkBM OH (Ω) ≤ C3 . Theorem 4.1 tells us that if a(u, Ω) ≥ 0 for all u ∈ W01,2 (Ω, R2 ), then kbkBM OH (Ω) ≤ C, that is, kbkBM Or (Ω) ≤ C by Theorem 3.1. From the Remark in [Z, page 426], if kbkBM Or (Ω) is sufficient small, then a(u, Ω) ≥ 0. We see that kbkBM Or (Ω) ≤ C is “almost” a necessary and sufficient condition of a(u, Ω) ≥ 0 for all u ∈ W01,2 (Ω, R2 ). Acknowledgement. The author would like to thank Professor Alan McIntosh for his supervision and for a meticulous reading of the manuscript. He also likes to thank the Centre for Mathematics and its Applications, the Australian National University and Australian Government for financial support. References [AR] [C] [CKS]
[CLMS]
P. Auscher, E. Russ, Hardy spaces and divergence operators on strongly Lipschitz domains of RN , J. Funct. Anal. to appear. D. C. Chang, The dual of Hardy spaces on a domain in RN , Forum Math. 6(1994), 65-81. D. C. Chang, S. G. Krantz, E. M. Stein, Hp theory on a smooth domain in RN and elliptic boundary value problems, J. Funct. Anal. 114(1993), 286-347. R. Coifman, P. L. Lions, Y. Meyer, S. Semmes, Compensated compactness and Hardy spaces, J. Math. Pures and Appl. 72(1993), 247-286.
JACOBIANS ON LIPSCHITZ DOMAINS OF R2
109
[GHL]
J. E. Gilbert, J. A. Hogan, J. D. Lakey, Atomic decomposition of divergence-free Hardy spaces, Mathematica Moraviza, Special Volume, 1997, Proc. IWAA, 33-52. [GR] V. Girault, P. A. Raviart, Finite Element Methods for Navier-Stokes Equations, theory and algorithms, Springer-Verlag, Berlin Heidelberg, 1986. [HLMZ] J. A. Hogan, C. Li, A. McIntosh, K. Zhang, Global higher integrability of Jacobians on bounded domains, Ann. Inst. H. Poincar´e Anal. Non lin´eaire 17(2000), 193-217. [J] P. W. Jones, Extension theorems for BM O, Indiana Univ. Math. J. 29(1980), 41-66. [JSW] A. Jonsson, P. Sj¨ ogren, H. Wallin, Hardy and Lipschitz spaces on subsets of RN , Studia Math. 80(1984), 141-166. [M] P. Marcellini, Quasiconvex quadratic forms in two dimensions, Appl. Math. Optim. 11(1984), 183-189. [N] J. Ne˘cas, Les m´ethodes directes en th´eorie des ´equations elliptiques. (French) Masson et Cie, Eds., Paris; Academia, Editeurs, Prague 1967. [RR] H. M. Riemann, T. Rychener, Funktionen Beschrankter Mittlerer Oszillation (German), Lecture Notes in Mathematics Vol. 487, Springer-Verlag, Berlin - New York, 1975. [S1] E. M. Stein, Singular Integrals and Differentiability Properties of Functions, Princeton Univ Press, Princeton, 1970. [S2] E. M. Stein, Harmonic Analysis, real-variable methods, orthogonality, and oscillatory integrals, Princeton Univ. Press, New Jersey, 1993. [Z] K. Zhang, On the coercivity of elliptic systems in two dimensional spaces, Bull. Austral. Math. Soc. 54(1996), 423-430 Department of Mathematics, Shantou University, Shantou Guangdong 515063 P. R. China; Centre for Mathematics and its Applications, Mathematical Sciences Institute, The Australian National University, Canberra ACT 0200 Australia E-mail address: [email protected]
DIVERGENT SUMS OF SPHERICAL HARMONICS CHRISTOPHER MEANEY Abstract. We combine the Cantor-Lebesgue Theorem and Uniform Boundedness Principle to prove a divergence result for Ces`aro and Bochner-Riesz means of spherical harmonic expansions.
1. Background Fix an integer d > 1 and consider the unit sphere S d in Rd+1 , equipped with normalized rotation-invariant measure. For each n ≥ 0 let Hn denote the space of spherical harmonics of degree n restricted to S d , so that L2 (S d ) = ⊕∞ n=0 Hn . See [22, Section 4.2] for details. Every d distribution ψ on S has a spherical harmonic expansion ∞ X (1) Yn (ψ)(x), ∀x ∈ S d , where Yn (ψ) ∈ Hn , ∀n ≥ 0. n=0
This is the expansion of ψ in eigenfunctions of the Laplace-Beltrami operator on S d . It is known [14] that if 1 ≤ p < 2 then there is an ψ ∈ Lp (S d ) for which (1) diverges almost everywhere. That leaves open the general behaviour of spherical harmonic expansions for elements of L2 (S d ). A partial step in this direction follows from the localization principle [18]. Theorem 1.1 (Localization). Suppose ψ is a distribution on S d and U ⊂ S d is an open P set disjoint from the support of ψ. For each x ∈ U , the expansion ∞ n=0 Yn (ψ)(x) converges if and only if Yn (ψ)(x) → 0 as n → ∞. Corollary 1.2. If ψ ∈ L2 (S d ) and U ⊂ S d is P an open set on which ψ is zero almost everywhere, then the expansion ∞ n=0 Yn (ψ)(x) converges to zero almost everywhere on U . There are special cases where a function ψ ∈ L2 (S d ) can be guaranteed to have an almost everywhere convergent spherical harmonic expansion, if ψ is in an L2 - Sobolev space W 2,s of positive index s [16] or if it is zonal [1]. (Recall that a function f on S d is said to be zonal about a point y ∈ S d when f (x) depends only on x · y for all x ∈ S d .) Date: 16 January 2002. 1991 Mathematics Subject Classification. 42C05,33C45,42C10. Key words and phrases. Jacobi polynomial, zonal function, Ces`aro mean, Riesz mean, Cantor-Lebesgue theorem, uniform boundedness. 110
DIVERGENT SUMS OF SPHERICAL HARMONICS
111
Carleson’s theorem [3] has been extended to zonal functions [11]. Let pc be the critical index 2d pc = . d+1 Theorem 1.3. If pc < p ≤ 2 and f ∈ Lp (S d ) is zonal about a point y ∈ S d , then its spherical harmonic expansion is convergent almost everywhere. Corollary 1.4. Suppose ψ ∈ L2 (S d ), U ⊂ S d is an open set, f1 ∈ S 2,s (S d ), f2 ∈ L2 (S d ) is a finite sumPof zonal functions, and s>0 W ψ = f1 + f2 almost everywhere on U . Then ∞ n=0 Yn (ψ)(x) converges almost everywhere on U . The two corollaries 1.2 and 1.4 would be rendered trivial if there where a higher dimensional version of Carleson’s theorem. They do suggest that when considering convergence of expansions, we should examine the term-wise behaviour away from the support of a distribution. In the early 1980’s we showed [17] that Theorem 1.3 is sharp and that localization fails at the critical index. Theorem 1.5. For each y ∈ S d and 1 ≤ p ≤ pc there is a ψ ∈ Lp (S d ), supported in the hemisphere {x : x · y ≥ 0} whose spherical harmonic expansion diverges almost everywhere. This was proved by a combination of the Cantor-Lebesgue theorem, 0 knowledge of the Lp -norms of the zonal spherical functions, and the uniform boundedness principle. Kanjin [13] showed that these methods could be combined with a result of Hardy and Riesz [12] to deal with Riesz means for radial functions on Euclidean space. This approach was also used in [20] for Riesz means of radial functions on non-compact rank one symmetric spaces. Here we prove a similar result for Ces`aro and Riesz means of spherical harmonic expansions of zonal functions. This shows the sharpness of the results in [4]. See [2, 7] for earlier work on Ces`aro means of spherical harmonic expansions. See [21, 5] for results in a more general setting. `ro & Riesz means 2. Cesa 2.1. Ces` aro means. The Ces`aro means [24, pages 76–77] of order δ of the expansion (1) are defined by (2)
δ σN ψ(x) =
N X AδN −n n=0
AδN
Yn (ψ)(x),
∀N ≥ 0, x ∈ S d ,
n+δ . Theorem 3.1.22 in [24] says that if the Ces`aro n means converge, then the terms of the series have controlled growth. where Aδn =
112
CHRISTOPHER MEANEY
δ ψ(x) exists for some x ∈ X and Lemma 2.1. Suppose that lim σN N →∞ δ > −1. Then |YN (ψ)(x)| ≤ Cδ N δ max σnδ ψ(x) , ∀n ≥ 0. 0≤n≤N
2.2. Riesz means. Hardy and Riesz [12] had proved a similar result for Riesz means. Recall that the Riesz means of order δ ≥ 0 are defined for each r > 0 by δ X k δ 1− (3) Sr ψ(x) = Yn (ψ)(x). r 0≤n 0 and x ∈ X at which its Riesz means Srδ ψ(x) converges to c as r → ∞ then 0 Sr ψ(x) − c ≤ Aδ rδ sup Stδ ψ(x) . 0
Note that this implies Yn (ψ)(x) = O(nδ ) and we have the same growth estimates as in Lemma 2.1. Gergen[9] wrote formulae relating the Riesz and Ces`aro means of order δ ≥ 0, from which it follows that the two methods of summation are equivalent. 3. Zonal Functions and Jacobi Polynomials 3.1. Notation. Suppose that f is a function on S d with f (x) depending only on x·y, for a fixed y ∈ S d , so that f (x) = f0 (x · y). The spherical harmonic expansion of f is ∞ X (α,α) (4) cn (f0 )h−1 (x · y) n Pn n=0 (α,α)
where α = (d − 2)/2, Pn index (α, α), Z hn =
1
is the Jacobi polynomial of degree n and
(α,α) 2 Pn (t) 1 − t2 α dt,
−1
and the coefficients are Z 1 α cn (f0 ) = f0 (t)Pn(α,α) (t) 1 − t2 dt,
∀n ≥ 0.
−1
See section 4.7 of Szeg˝o’s book [23] for details about these special functions. Let mα be the measure on [−1, 1] given by dmα (t) = (1 − t2 )α dt,
DIVERGENT SUMS OF SPHERICAL HARMONICS
113
(α,α)
so that {Pn : n ≥ 0} is a orthogonal basis of L2 (mα ). From (4.3.3) in [23] we know that the normalization constants hn satisfy (5)
hn ∼ A n−1 as n → ∞
3.2. Uniform Boundedness. Suppose there is a number 1 < q ≤ ∞ and some positive number A with kPn(α,α) kLq (mα ) ≥ cnA , The formation of the coefficient Z F 7→ cn (F ) =
∀n ≥ 1.
1
F (t)Pn(α,α) (t)dmα (t)
−1
is then a bounded linear functional on the dual of Lq (mα ) with norm bounded below by a constant multiple of nA . The uniform boundedness principle implies that for p conjugate to q and each 0 ≤ ε < A there is an F ∈ Lp (mα ) so that (6)
cn (F )/nε → ∞ as n → ∞.
3.3. Cantor-Lebesgue Theorem. This idea is explained in [19] and is based on [24, Section IX.1]. Suppose we have a sequence of functions Fn on an interval in the real line with the asymptotic property Fn (θ) = cn (cos(Mn θ + γn ) + o(1)) ,
∀n ≥ 0
uniformly on a set E of finite positive measure, and with Mn → ∞ as n → ∞. Integrating |Fn |2 over E gives Z Z 2 2 2 |Fn (θ)| dθ = |cn | cos (Mn θ + γn ) dθ + o(1) E E |E| e2iγn e−2iγn 2 = |cn | + χ cE (2Mn ) + χ cE (−2Mn ) + o(1) . 2 4 4 The Riemann-Lebesgue Theorem [24, Thm. II.4.4] says that the Fourier transforms χ cE (±2Mn ) → 0 as Mn → ∞. If we know that there is some function G for which |Fn (θ)| ≤ G(n) uniformly on E for all n then there is an n0 > 0 for which Z |E| 2 |Fn (θ)|2 dθ ≤ G(n)2 |E|, ∀n ≥ n0 . |cn | ≤ 4 E This shows that |cn | ≤ 2G(n) for all n ≥ n0 . 3.4. Asymptotics. Theorem 8.21.8 in Szeg˝o’s book[23] gives the fol(α,α) lowing asymptotic behaviour for the Jacobi polynomials Pn . For α ≥ −1/2 and ε > 0 the following estimate holds uniformly for all ε ≤ θ ≤ π − ε and n ≥ 1. (7) Pn(α,α) (cos θ) = n−1/2 k(θ) cos (Mn θ + γ) + O n−3/2 . Here k(θ) = π −1/2 (sin(θ)/2)−α−1/2 , Mn = n + (2α + 1)/2, and γ = −(α + 1/2)π/2.
114
CHRISTOPHER MEANEY
From Egoroff’s theorem and Lemma 2.1 we can say that if the series (4) is Ces`aro summable of order δ on a set of positive measure in S d then there is a set of positive measure E ⊂ [0, π] on which (α,α) cn (f0 )h−1 ≤ Anδ P (cos θ) n n and hence (8)
cn (f0 )n(1/2)−δ cos (Mn θ + γ) + O(n−1 ) ≤ A
uniformly for θ ∈ E. The argument of subsection 3.3 shows that cn (f0 )n(1/2)−δ ≤ A, (9) ∀n ≥ 1. Lemma 3.1. If f is a zonal function on the unit sphere whose spherical harmonic expansion is Ces`aro summable of order δ on a set of positive measure, then there is a constant A > 0 for which |cn (f0 )| ≤ Anδ−(1/2) ,
∀n ≥ 1.
3.5. Norm Estimates. Markett[15] has calculated estimates on the Lp norms of Jacobi polynomials. Let qc =
4(α + 1) 2d = . 2α + 1 d−1
Equation (2.2) in [15] gives the following lower bounds on these norms. Lemma 3.2. For real number α > −1/2, 1 ≤ q < ∞, and r > −1/q, −1/2 Z 1 1/q if q < qc , n (α,α) q 1/q −1/2 Pn (x) (1 − x)α dx ∼ n (log n) if q = qc , 0 nα−(2α+2)/q if q > qc . Notice that these integrals are taken over [0, 1] rather than all of [−1, 1]. 4. Main Result Theorem 4.1. For each 1 ≤ p < pc = 2d/(d + 1), 0≤δ<
d d+1 − , p 2
and y ∈ S d , there is a function in Lp (S d ) which is zonal about y, supported in the hemisphere {x : x · y ≥ 0}, and whose spherical harmonic expansion has Ces`aro and Riesz means which diverge almost everywhere. Proof. Suppose that a series (4) has Ces`aro means of order δ which converge on a set of positive measure. Then Lemma 3.1 implies that (10) cn (f0 ) = O nδ−(1/2) , as n → ∞.
DIVERGENT SUMS OF SPHERICAL HARMONICS
115
Compare this inequality with the last line of Lemma 3.2 and section 3.2. If q > qc , (1/p) + (1/q) = 1 and α−
(2α + 2) 1 >δ− q 2
then there must be a zonal function f ∈ Lp (S d ) with f0 supported on [0, 1] for which the estimate (10) fails. Remembering the definition of α in terms of the dimension d, we are considering 1 d−1 1 δ− < −d 1− 2 2 p which means δ<
d (d + 1) − . p 2
Remark 4.1. In [19] we applied this technique to produce an analogous theorem for Laguerre expansions. 5. Central function on SU (2) We conclude with a simple three dimensional example. Suppose that G = SU (2) is equipped with the normalized translation invariant measure µ and that T is the maximal torus of diagonal elements of G. b = {k/2 : k ∈ Z, k ≥ 0} there is an irreducible unitary For each ` ∈ G representation of G with dimension 2` + 1 and character iθ sin ((2` + 1)θ) e 0 . χ` = 0 e−iθ sin (θ) Every central function on G is determined by its restriction to T . The Fourier series of central functions are expansions in the characters. If f ∈ L1 (G, µ) is central then ∞ X (11) f∼ c` χ` `=0
with Z (12)
c` =
f (x)χ` (x) dµ(x),
b ∀` ∈ G.
G
In [8] and [10], Dooley, Giulini, Soardi, and Travaglini estimated the Lebesgue norms of characters of compact Lie groups. The group SU (2) provides the simplest case of these estimates. For each q > 3 b (13) kχ` k ≥ c(2` + 1)1−3/q , ∀` ∈ G. q
If 1/p + 1/q = 1 then 3 1 3 1− =1−3 1− = − 2. q p p
116
CHRISTOPHER MEANEY
Uniform boundedness then says that if 1 ≤ p < 3/2 and a < (3/p) − 2 then there is a central function f ∈ Lp (G) for which the coefficients in (11) have c` /(2` + 1)a unbounded as ` → ∞. Suppose that (11) is Ces`aro summable of order δ on a set of positive measure. Then Lemma 2.1 says that c` sin ((2` + 1)θ) = O(`δ ) as ` → ∞, on a set of positive measure. The Cantor-Lebesgue Theorem then says that c` = O(`δ ) as ` → ∞. Theorem 5.1. For 1 ≤ p < 3/2 and 0 ≤ δ < (3/p) − 2 there is a central function f ∈ Lp (SU (2)) for which the Ces` aro and Riesz means of order δ are divergent almost everywhere. This shows the sharpness of results in Clerc’s paper [6]. References 1. V. M. Badkov, Approximation properties of Fourier series in orthogonal polynomials., Russ. Math. Surv. 33 (1978), no. 4, 53–117 (English). 2. Aline Bonami and Jean-Louis Clerc, Sommes de Ces` aro et multiplicateurs des d´eveloppements en harmoniques sph´eriques, Trans. Amer. Math. Soc. 183 (1973), 223–263. 3. Lennart Carleson, On convergence and growth of partial sumas of Fourier series, Acta Math. 116 (1966), 135–157. 4. Sagun Chanillo and Benjamin Muckenhoupt, Weak type estimates for Ces` aro sums of Jacobi polynomial series, Mem. Amer. Math. Soc. 102 (1993), no. 487, viii+90. 5. F. M. Christ and Christopher D. Sogge, The weak type L1 convergence of eigenfunction expansions for pseudodifferential operators, Invent. Math. 94 (1988), no. 2, 421–453. 6. Jean-Louis Clerc, Sommes de Riesz et multiplicateurs sur un groupe de Lie compact, Ann. Inst. Fourier (Grenoble) 24 (1974), no. 1, 149–172. 7. Leonardo Colzani, Mitchell H. Taibleson, and Guido Weiss, Maximal estimates for Ces` aro and Riesz means on spheres, Indiana Univ. Math. J. 33 (1984), no. 6, 873–889. 8. A. H. Dooley, Norms of characters and lacunarity for compact Lie groups, J. Funct. Anal. 32 (1979), no. 2, 254–267. 9. J.J. Gergen, Summability of double Fourier series., Duke Math. J. 3 (1937), 133–148 (English). 10. Saverio Giulini, Paolo M. Soardi, and Giancarlo Travaglini, Norms of characters and Fourier series on compact Lie groups, J. Funct. Anal. 46 (1982), no. 1, 88–101. 11. J. J. Guadalupe, M. P´erez, F. J. Ruiz, and J. L. Varona, Two notes on convergence and divergence a.e. of Fourier series with respect to some orthogonal systems, Proc. Amer. Math. Soc. 116 (1992), no. 2, 457–464. 12. G. H. Hardy and M. Riesz, A general theory of Dirichlet series, 1915. 13. Y¯ uichi Kanjin, Convergence almost everywhere of Bochner-Riesz means for radial functions, Ann. Sci. Kanazawa Univ. 25 (1988), 11–15.
DIVERGENT SUMS OF SPHERICAL HARMONICS
117
14. Carlos E. Kenig, Robert J. Stanton, and Peter A. Tomas, Divergence of eigenfunction expansions, J. Funct. Anal. 46 (1982), no. 1, 28–44. 15. C. Markett, Cohen type inequalities for Jacobi, Laguerre and Hermite expansions, SIAM J. Math. Anal. 14 (1983), no. 4, 819–833. 16. Christopher Meaney, On almost-everywhere convergent eigenfunction expansions of the Laplace-Beltrami operator, Math. Proc. Cambridge Philos. Soc. 92 (1982), no. 1, 129–131. 17. , Divergent Jacobi polynomial series, Proc. Amer. Math. Soc. 87 (1983), no. 3, 459–462. , Localization of spherical harmonic expansions, Monatsh. Math. 98 18. (1984), no. 1, 65–74. 19. , Divergent Ces` aro and Riesz means of Jacobi and Laguerre expansions, Preprint http://xxx.lanl.gov/abs/math.CA/0202019, Feb. 2002. To appear in Proc. Amer. Math. Soc. 20. Christopher Meaney and Elena Prestini, Bochner-Riesz means on symmetric spaces, Tohoku Math. J. (2) 50 (1998), no. 4, 557–570. 21. Christopher D. Sogge, On the convergence of Riesz means on compact manifolds, Ann. of Math. (2) 126 (1987), no. 2, 439–447. 22. Elias M. Stein and Guido Weiss, Introduction to Fourier analysis on Euclidean spaces, Princeton University Press, Princeton, N.J., 1971, Princeton Mathematical Series, No. 32. 23. G´ abor Szeg˝ o, Orthogonal polynomials, third ed., American Mathematical Society, Providence, R.I., 1967, American Mathematical Society Colloquium Publications, Vol. 23. 24. A. Zygmund, Trigonometric series: Vols. I, II, Cambridge University Press, London, 1968. Dept. Mathematics, Macquarie University, North Ryde, NSW 2109 Australia E-mail address: [email protected]
DERIVATION OF MONOGENIC FUNCTIONS AND APPLICATIONS T. QIAN Abstract. The paper studies two types of results on inducing monogenic functions in Rn1 . One is based on McIntosh’s formula and the other is along the line of Fueter’s Theorem. Applications are summarized and a new application on monogenic sinc function interpolation is introduced.
1. Background Denote by e1 , e2 , . . . , en the basic elements that satisfy e2i = −1, ei ej = −ej ei , i, j = 1, 2, . . . , n, i < j. We will work on the following spaces: Rn = {x = x1 e1 + · · · xn en : xi ∈ R, i = 1, . . . , n}, Rn1 = {x = x0 + x : x0 ∈ R, x ∈ Rn }, R(n) is the Clifford algebra generated by e1 , e2 , . . . , en over the real number field R; C(n) is the Clifford algebra generated by e1 , e2 , . . . , en over the complex number field C. P We adopt the notation x ∈ R(n) (or C(n) ) implies x = s xs es , where xs ∈ R (or C), and s runs over all the possible ordered sets s = {0 ≤ j1 < · · · < jk ≤ n}, or s = ∅, es = ej1 · · · ejk ,
and
e0 = e∅ = 1.
The functions we will study will be defined in subsets of Rn1 , and take their values in R(n) or C(n) . 1
1991 Mathematics Subject Classification. Primary: 42A50, 30G35; Secondary: 42B30, 30A05 2 Key Words and Phrases: Dirac Operator, Monogenic Functions, Clifford Algebra, Fourier Analysis. 3 The study was supported by Research Grant of the University of Macau No. RG024/00-01S/QT/FST. 118
DERIVATION OF MONOGENIC FUNCTIONS
119
The Dirac operator D for functions in Rn1 is defined by ∂ ∂ ∂ , D= e1 + · · · + en . ∂x0 ∂x1 ∂xn It applies from the left- and right- hand sides to the function, in the manners n X n X X X ∂fs ∂fs ei es and f D = es ei , Df = ∂xi ∂xi i=0 s i=0 s D = D0 + D, D0 =
respectively. If Df = 0, then f is said to be left-monogenic; and, if f D = 0, then right-monogenic. If f is both left- and right- monogenic, then it is said to be monogenic. Examples: (1) The case n = 1 corresponds to the complex number field: e1 = i, ∂ ∂ + i ∂y , f (z) = u(x, y) + iv(x, y) and Df = 0 if and only if the D = ∂x Cauchy-Riemann equations hold: ∂u ∂v = ∂x ∂y ∂u ∂v =− . ∂x ∂y (2) The case n = 2 corresponds to the space of Hamilton quaternions: ~i = e1 ,~j = e2 , ~k = e1 e2 , q = q0 + q1~i + q2~j + q3~k, qk ∈ R. Profound studies on Clifford analysis have been conducted since Fueter’s school in the 1930’s till the present time (see, for instance, [Ma], [CS] and [Q1] and their references). (3) Let uj (x), j = 0, 1, . . . , n, be defined in Rn1 with values in C. Set U = −u0 + u1 e1 + · · · + un en Then DU = 0 if and only if these functions form a conjugate harmonic system (or satisfy the generalized Cauchy-Riemann equations, see [St] and [KQ1]): n X ∂uj =0 ∂xj j=0 ∂uk ∂uj = , 0 ≤ k < j ≤ n. ∂xj ∂xk
120
T. QIAN
For monogenic functions there hold Cauchy’s Theorem and Cauchy’s x formula. The Cauchy kernel in the context is E(x) = |x|n+1 , where for x = x0 + x, we denote x = x0 − x .. It is observed that the Clifford structure of Rn1 is the “true” analogue of the one complex variable structure of R11 . 2. C-K Extension and McIntosh’s Formula It can be proved that if we have a real analytic function defined in an open set O of Rn , then we can always monogenically extend it to an open set Q of Rn1 where O = Rn ∩ Q (C-K extension, see, for instance [BDS]). The extension can be realized by the operation e−x0 D f (x), understood in the symbolic way. In fact, formally we have D(e−x0 D f (x)) = (D0 + D)(e−x0 D f (x)) = (−D)e−x0 D f (x) + De−x0 D f (x) = 0. Examples: (1) If f (x) = xj , then 1 4 (−x0 D)2 + · · · )xj = xj e0 − x0 ej = zj . 2! (2) The extension of xi xj , i 6= j, is 1 (zi zj + zj zi ), 2 etc. In practice the C-K extension and the related forms are, in general, complicated and not easy to use. On the contrary, McIntosh’s formula, somehow plays the role of Fourier-Laplace transform in Rn1 , has been playing a crucial role in a number of questions in function theory ([LMcQ], [PQ], [Q4], [KQ1], [KQ2]). The formula first appeared in late 1980’s ([Mc1]) and formally published in [Mc2] and [LMcQ] in 1994. The formula involves a set of notations: If f is defined in Rn with Fourier transform, then the possible monogenic extension of f is given by Z ∧ 1 f (x) = e(x, ξ) f (ξ)dξ, (McIntosh’s formula), (2π)2 Rn e−x0 D (xj ) = (1 + (−x0 D) +
provided that the integral on the right-hand-side is properly defined, where e(x, ξ) = eix·ξ {e−x0 |ξ| χ+ (ξ) + ex0 |ξ| χ− (ξ)}, ξ 1 χ± (ξ) = (1 + i ), 2 |ξ|
x = x0 + x, x, ξ ∈ Rn .
DERIVATION OF MONOGENIC FUNCTIONS
121
In [BDS] a wide range of similar notions are introduced. It is exactly McIntosh’s form, however, that has been effectively used, especially in problems related to Fourier transformation. In the formulas for the projections χ± , if we take n = 1 and e1 = −i, then we have χ± (ξ) = ±sgnξ, where ξ = iξ. This indicates that the formula provides a decomposition of a function into functions similar to those in the Hardy spaces. Indeed, we have, f (x) = f + (x) + f − (x), where Z ∧ 1 4 ± ± e (x, ξ) f (ξ)dξ, f (x) = (2π)n Rn e± (x, ξ) = eix·ξ e∓x0 |ξ| χ± (ξ). We can further show that for x0 > 0, Z 1 + f (x) = E(x − y) · f (y)dy; wn Rn while f − (x), x0 > 0, is the monogenic extension of f − (x) for x0 < 0, where the latter is also of the Cauchy’s integral form of f . For x0 < 0 we have the analogous notation. Under the context of the classical Paley-Wiener Theorem in the case n = 1, viz. f ∈ L2 (R) and ∧
supp f (ξ) ⊂ [−δ, δ],
δ > 0,
there follows f (z) = f + (z) + f − (z). For z = x + iy, y > 0, we have e−y|ξ| χ[0,δ] (ξ) = e−yξ χ[0,δ] (ξ),
ey|ξ| χ[−δ,0] (ξ) = e−yξ χ[−δ,0] (ξ),
and so 1 f (z) = 2π +
Z
and
δ
e 0
ixξ −yξ
e
∧
1 f (ξ)dξ = 2πi
Z
∞
−∞
f (t) dt, t−z
Z 0 ∧ 1 eixξ e−yξ f (ξ)dξ, f (z) = 2π −δ − where f (z) is well defined, but not be expressible by a Cauchy integral. In fact, since y = Im z > 0, f − (z) is the holomorphic extension to the upper-half complex plane of the Cauchy integral of f in the lower-half −
122
T. QIAN
complex plane. By virtue of McIntosh’s formula we have exactly the same notion in Rn1 . We will mention two applications of McIntosh’s formula. (1) Paley-Wiener Theorem in Rn1 In a recent paper we proved the following theorem ([KQ1]). Theorem. Let f ∈ L2 (Rn ). Then f can be monogenically extended to Rn1 with the estimate |f (x)| ≤ ceR|x| if and only if ∧
supp f ⊂ B(0, R), where B(0, R) = {x ∈ Rn : |x| < R}. In the case we have Z 1 e(x, ξ)fˆ(ξ)dξ, x ∈ Rn1 . f (x) = (2π)n Rn In the literature higher dimensional versions of the Paley-Wiener theorem have been sought (see the references of [KQ1]). We wish to make the point that the version with the Clifford algebra setting provides the precise analogue. The commonly adopted proofs of the classical Paley-Wiener Theorem are not readily applicable to the Clifford setting owing to the defect that products of monogenic functions are no longer again monogenic in general. However, a particular proof for the one complex variable case can be closely followed through a non-trivial computation based on McIntosh’s formula([KQ1]). (2) Monogenic sinc function with Shannon sampling for functions in the Paley-Wiener classes Define the class of functions P W (R) = f : Rn1 → C(n) : f is monogenic in the whole Rn1 and satisfies |f (x)| ≤ CeR|x| . The monogenic sinc function is defined to be Z 1 sinc(x) = e(x, ξ)dξ. (2π)n [−π,π]n The following exact interpolation of functions in the P W (R) classes is proved in ([KQ2]). Theorem. If f ∈ P W ( πh ) , then X x − hk f (x) = f (hk) sinc , h k∈Zn
DERIVATION OF MONOGENIC FUNCTIONS
123
where the convergence is in the pointwise sense independent of the order of summation. The proof is based on estimates of the monogenic sinc function derived from McIntosh’s formula. 3. Fueter’s Theorem and Generalizations This addresses the problem of deriving monogenic and harmonic functions from those of the same kind but in lower dimensional spaces. Let f 0 (z) be a function of one complex variable analytic in an open set O of the upper-half complex plane C+ . If f (z) = u(x, y) + iv(x, y), z = x + iy, we introduce x f~0 (x) = u(x0 , |x|) + v(x0 , |x|). |x| and set n−1 τ (f 0 )(x) = ∆ 2 f~0 (x), x ∈ Rn1 . Theorem of Fueter (1935). When n = 3, interpreted as the quaternionic space, the mapping τ maps an analytic function f 0 (z) in O to a quaternionic monogenic function in ~ = {q = q0 + q : q0 + i|q| ∈ O}. O Theorem of Sce (1957). For n being an odd integer the mapping τ maps f 0 (z) to a monogenic function in ~ = {x = x0 + x : x0 + i|x| ∈ O}. O These results were extended in [Q2] in 1997 to the cases n being n−1 an integer and the operator ∆ 2 interpreted as the Fourier multiplier operator with symbol |ξ|n−1 . We note that 1 1 τ ( )(x) = E(x) = n +1 . z |x| In [Q2-3] for any integer n ≥ 2 a corresponding relationship between the functions f 0 (z) = z k and certain monogenic functions P (k) (x) of homogeneity of degree k is established: 1 τ ( k )(x) = P (−k) (x), k = 1, 2, . . . , z and P (k−1) = I(P (−k) ), k = 1, 2, . . . , where I is the Kelvin inversion defined by If (x) = E(x)f (x−1 ).
124
T. QIAN
It is noted that if n is an odd integer, then P (k−1) = τ (z n+k−2 ). The sequence P (k) , k ∈ Z, is used to establish the bounded holomorphic functional calculus of the Dirac operator on Lipschitz perturbations,denoted by DΣ , of the unit sphere in Rn1 (and similarly on Rn ). We now describe the result. Set |y| π Sw = {0 6= z ∈ C : z = x + iy, < tan w}, 0 < w < , |x| 2 tan w > the Lipschitz constant of Σ, ∞ H (Sw ) = {b : Sw → C : f o is bounded and analytic in Sw }. Given b ∈ H∞ (Sw ), and set, formally, Z 1 b(DΣ )f = b(ξ)(I − DΣ )−1 dξf, 2πi γ where γ is a certain curve in Sw surrounding the spectrum of DΣ . The operators b(DΣ ) are proved to be equal to the Fourier multiplier operators ∞ ∞ X X Mb f (x) = b(k)Pk f (x) + b(−k)Qk f (x), k=1
k=1
where Pk f and Qk f are projections of f onto the spaces of monogenic functions of homogeneity degree k and −k , respectively. They are also equal to the singular integral operators Z 1 −1 1 SΦ f (x) = lim Φ(y x)E(y)n(y)f (y)dσ(y) + Φ (, x)f (x) , n→∞ wn |x−y|> ∨
where in a certain sense Φ =b (the inverse Fourier transform of b) and Φ1 (, x) = the average of Φ on the sphere centered at x of radius . That is b(DP ) = Mb = Sφ . The boundedness of the operators b(DΣ ), b ∈ H∞ (Sw ), is proved through their singular integral expressions SΦ based on the estimates of the kernels Φ and Φ1 . The derivation of the estimates are reduced to the similar estimates in the one complex variable case via the correspondence between the functions z k and P (k) ([Q5]). On Lipschitz perturbations of higher dimensional spheres the theory cannot be done through the Poisson Summation method based on the graph case as in the unit circle case ([Q5]). It encountered some difficulties and hence was first achieved in the quaternionic space ([Q1]), and then in general Euclidean spaces ([Q3]).
DERIVATION OF MONOGENIC FUNCTIONS
125
Further generalizations of Fueter’s Theorem include the following. (i) In a recent paper F.Sommen proved that if n is an odd positive integer and x ∈ Rn1 , then for f 0 (z) = u(s, t)+ iv(s, t), z = s + it, ~ analytic in an open set O ⊂ C+ , then for x ∈ O n−1 x D∆k+ 2 ((u(x0 , |x|) + v(x0 , |x|))Pk (x)) = 0, |x| where Pk is any polynomial in x of homogeneity k, left-monogenic with respect to the Dirac operator D ([So]). (ii) K.I.Kou and T.Qian extended Sommen’s result to the cases when n is an even positive integer and Sommen extended his being non-negative integers, no matter result to the cases k+ n−1 2 whether k is an integer ([KQS]). (iii) The derivation of monogenic functions can be reduced to that of harmonic functions, based on the following observations. A. If h is harmonic in x0 , x1 , . . . , xn , then Dh is monogenic, where D = D0 − D. B. If f is monogenic, then there exists a harmonic function h such that f = Dh. The following result for harmonic functions is obtained in a recent paper of T.Qian and F.Sommen ([QS]). Denote (r) (r)
(r) pr x(r) = x1 e1 + · · · + x(r) pr epr ∈ R , P where r = 1, . . . , d, dr=1 pr = m, and (r) (r 0 )
ei ei0
(r 0 ) (r)
= −ei0 ei ,wherever (r, i) 6= (r0 , i0 ).
Let h(s1 , . . . , sd ) be a harmonic function in the d variables P s1 , . . . , sd . Them, if pr , r = 1, . . . , d, are odd and m = dr=1 pr is even, then m
∆ 2 h(|x(1) |, . . . , |x(d) |) = 0, where ∆ is the Laplacian for all the m variables xri , r = 1, . . . , d, i = 1, . . . , pr . (iv) The latest result along this line is by K.I.Kou and T.Qian ([KQ3]), as follows. In the above notation we have m
(1)
(d)
∆(k1 +...+kd )+ 2 [h(|x(1) |, . . . , |x(d) |)Pk1 · · · Pkd (x(d) )] = 0, (r)
where for any r = 1, 2, . . . , d, Pkr (x(r) ) is a left-monogenic functions with respect to D(r) , homogeneous of degree kr , where kr is any non-negative integer.
126
T. QIAN
References [BDS] F. Brackx, R. Delanghe and F. Sommen, Clifford Analysis, Research Notes in Mathematics, Vol. 76, Pitman Advanced Publishing Company, Boston, London, Melbourne (1982). [F] R. Fueter, Die Funktionentheorie der Differetialgleichungen ∆u = 0 und ∆∆u = 0 mit vier reellen Variablen, Comm. Math. Helv. 7(1935), 307-330. [GS] K. G¨ urlebeck and W. Spr¨ossig, Quaternionic analysis and elliptic boundary value problems, Birkh¨ auser, Basel. [KQ1] K.I. Kou and T. Qian, The Paley-Wiener Theorem in Rn with the Clifford analysis setting, J. of Func. Anal., 189, 227-241 (2002). [KQ2] K.I. Kou and T. Qian, Shannon sampling using the monogenic sinc finction, preprint. [KQ3] K.I. Kou and T. Qian, A Note on Derivation of harmonic functions in higher dimensional spaces, preprint. [KQS] K.I. Kou, T. Qian and F. Sommen, Generalizations of Fueter’s Theorem, preprint. [LMcQ] C. Li, A. McIntosh and T. Qian, Clifford algebras, Fourier transforms, and singular convolution operators on Lipschitz surfaces, Revista Matem´atica Iberoamericana, Vol. 10, No.3, (1994), 665-721. [Mc1] A. McIntosh, manuscripts in Clifford analysis, private communication, 1986. [Mc2] A. McIntosh, Clifford algebras, Fourier theory, singular integrals and harmonic functions on Lipschitz domains, Studies in Advanced Mathematics: Clifford Algebras in Analysis and Related Topics, edited by John Ryan, CRC Press, 1996. [Ma] H.R. Malonek, Rudolf Fueter and his motivation for hypercomplex function theory, preprint. [PQ] J. Peetre and T. Qian, M¨ obius covariance of iterated Dirac operators, J. Austral. Math. Soc., Series A, 56 (1994), 1-12. [Q1] T. Qian, Singular integrals on star-shaped Lipschitz surfaces in the quaternionic space, Math. Ann., 310 (4) (April, 1998), 601-630. [Q2] T. Qian, Generalization of Fueter’s result to Rn+1 , Rend. Mat. Acc. Lincei, s.9, v.8 (1997), 111-117. [Q3] T. Qian, Fourier analysis on star-shaped Lipschitz surfaces, J. of Func. Anal. 183, 370-412 (2001). [Q4] T. Qian, Singular integrals with Monogenic kernels on the m-torus and their Lipschitz perturbations, Studies in Advanced Mathematics: Clifford Algebras in Analysis and Related Topics, edited by John Ryan, CRC Press, 1996. [Q5] T. Qian, Singular integrals with holomorphic kernels and H ∞ -Fourier multipliers on star-shaped Lipschitz curves, Studia mathematica, 123 (3) (1997), 195-216. [QS] T. Qian and F. Sommen, Deriving harmonic functions in higher dimensional spaces, preprint. [Sc] M. Sce, Osservazioni sulle serie di potenze nei moduli quadratici, Atti Acc. Lincei Rend. fis., s.8, 23, 1957, 220-225. [So] F. Sommen, On a generalization of Fueter’s theorem, Zeitschrift f¨ ur Analysis und ihre Anwendungen, Journal for Analysis and its Applications, Vol 19 (2000), N0.4, 899-902.
DERIVATION OF MONOGENIC FUNCTIONS
127
[St]
E.M. Stein, Singular integrals and differentiability properties of functions, Princeton University Press, 1970. [SVL] F. Sommen and P. Van Lancker, Homogeneous monogenic functions in Euclidean spaces, Integral Transforms and Special Functions, 1998, Vol 7, No. 3-4, 285-298. [VL1] P. Van Lancker, Clifford Analysis on the Unit Sphere, Ph.D. thesis, University of Gent, 1996. [VL2] P. Van Lancker, Clifford Analysis on the Sphere, V. Dietrich et al. (eds.), Clifford Algebras and Their Applications in Mathematical Physics, Kluwer Acad. Publ., 1998, 201-215. Faculty of Science and Technology, University of Macau, P. O. Box 3001 Macau E-mail address: [email protected]
ASYMPTOTICS OF SEMIGROUP KERNELS A.F.M. TER ELST AND DEREK W. ROBINSON Dedicated to Alan McIntosh on occasion of his 60-th birthday Abstract. We review the large time behaviour of the semigroup kernel associated with a homogeneous operator on a Lie group with polynomial growth. We consider complex second-order operators and two classes of higher order operators.
1. Introduction There is a vast literature on second-order elliptic operators on Lie groups with real symmetric coefficients. (See [Rob], [VSC], and references cited therein.) The closure of such an operator generates a semigroup which is holomorphic in a sector, the semigroup has a smooth kernel and the kernel, together with all its derivatives, satisfies the canonical Gaussian upper bounds for small time. These results have been extended to various other classes of complex subelliptic operators of any order on a Lie group and in particular one has again the Gaussian upper bounds for small time. If the operator is a real symmetric pure second-order operator and the Lie group has polynomial growth then the kernel satisfies the Gaussian upper bounds for all time. The aim of this note is to indicate the difficulties that one can expect for the large time canonical Gaussian upper bounds associated to other classes of operators on Lie groups with polynomial growth. In particular we discuss the class of pure second-order operators with complex constant coefficients and two classes of higher order homogeneous operators. Let G be a connected Lie group with Lie algebra g. Let a1 , . . . , ad0 be an algebraic basis for g, i.e., independent elements which together with their multi-commutators up to order s span g. The smallest number s for which this is valid is called the rank of the algebraic basis. Let dg be a (left) Haar measure on G. For all p ∈ [1, ∞] let L denote the left regular representation in Lp (G) = Lp (G ; dg). For all i ∈ {1, . . . , d0 } let Ai be the infinitesimal generator of the one parameter group t 7→ L(exp(−tai )). We also need multi-index since S∞ notation 0 0 n the Ai do not commute in general. Set J(d ) = n=0 {1, . . . , d } and 1991 Mathematics Subject Classification. 22E25, 35B40, 35J30. 128
ASYMPTOTICS OF SEMIGROUP KERNELS
129
if α = (i1 , . . . , in ) ∈ J(d0 ) set Aα = Ai1 . . . Ain and |α| = n. Associated to the algebraic basis there is a modulus on G, i.e., the distance to the identity element e of G. Introduce D as the set of functions 0
W = {ψ ∈
Cc∞ (G)
: ψ is real and sup
d X
g∈G
|(Ak ψ)(g)|2 ≤ 1} .
k=1
For all g ∈ G define |g|0 = sup |ψ(g) − ψ(e)| . ψ∈W
Finally, for all ρ > 0 let the volume V 0 (ρ) be the Haar measure of the ball {g ∈ G : |g|0 < ρ}. The Lie algebra g is called nilpotent if there exists an n ∈ N such that [b1 , [b2 , . . . , [bn−1 , bn ] . . .]] = 0 for all b1 , . . . , bn ∈ g. If g is nilpotent then define the rank r of g by r = max{n ∈ N : ∃b1 ,...,bn ∈g [b1 , [b2 , . . . , [bn−1 , bn ] . . .]] 6= 0} . Let g˜(d0 , r) be the nilpotent Lie algebra with maximal dimension, d0 generators and rank r. We denote the generators by a ˜1 , . . . , a ˜d0 and the e e e associated infinitesimal generators by A1 , . . . , Ad0 . Let G(d0 , r) be the connected simply connected Lie group with Lie algebra g˜(d0 , r). Now we are able to introduce the operators we want to consider. Let m ∈ 2N. On a general Lie group G the operator X H= c α Aα |α|≤m
T with cα ∈ C and domain D(H) = |α|≤m D(Aα ) is called subcoercive of step r if the comparable operator X e= eα H cα A |α|=m
e 0 , r)) satisfies a G˚ on L2 (G(d arding inequality, i.e., X eα ϕk e ϕ) kA ˜ 2˜2 Re(ϕ, ˜ H ˜ ≥µ ˜ |α|=m/2
e 0 , r)), for some µ ˜ > 0. Note that the condition is for all ϕ˜ ∈ Cc∞ (G(d on the principal part of the operator and is independent of the lower order terms. If H is subcoercive of step r then H is also subcoercive of step r − 1 (see [ElR1], Corollary 3.6).
130
A.F.M. TER ELST AND DEREK W. ROBINSON
Example 1.1. Let cij ∈ C and suppose there exists a µ > 0 such that 0
(1)
Re
d X
cij ξi ξj ≥ µ |ξ|2
i,j=1
P0 0 for all ξ ∈ Cd . Then the operator − di,j=1 cij Ai Aj is a subcoercive operator of step r for all r ∈ N. This is easy to verify. Conversely, P0 if − di,j=1 cij Ai Aj is a subcoercive operator of step r for some r ≥ 2 then (1) is valid. (See [ElR1], Proposition 3.7.) P0 m/2 P0 For all m ∈ 2N the operators − di=1 Ai 2 and (−1)m/2 di=1 Am i are also subcoercive operators of step r for all r ∈ N, but the proof is not trivial. (See [ElR2], Example 4.4.) e 2) is the Heisenberg group and the operator −A e1 2 − The group G(2, e2 2−10i[A e1 , A e2 ] is subcoercive of step 1, but not subcoercive of order r A for any r ≥ 2. The small time theory is well developed. Set 0 m t−1 )1/(m−1)
(m)
Gb,t (g) = V 0 (t)−1/m e−b((|g| ) for all b > 0, t > 0 and g ∈ G.
Theorem 1.2. Let a1 , . . . , ad0 be an algebraic basis of rank r for the Lie algebra of a Lie group G. Let m ∈ 2N and H an m-th order subcoercive operator of step s with s ≥ r. Then one has the following. I. The closure of H generates a semigroup S on Lp for all p ∈ [1, ∞]. II. The semigroup S is holomorphic. III. The semigroup has a smooth rapidly decreasing p-independent kernel K, i.e., St ϕ = Kt ∗ ϕ for all ϕ ∈ Lp and t > 0. IV. For all α ∈ J(d0 ) there exist b, c > 0 and ω ≥ 0 such that (2) V.
(m)
|Aα Kt | ≤ c t−|α|/m eωt Gb,t
for all t > 0. If p ∈ h1, ∞i then H is closed on Lp . Moreover, if the semigroup S is uniformly bounded on Lp then for all N ∈ N one has T N/m D(H ) = |α|=N D(Aα ).
Proof. See [ElR1], Theorems 2.5 and 4.1, and for Statement V see [BER]. The bounds (2) are optimal for t ∈ h0, 1] and describe the Gaussian decay. For large t the factor eωt reflects the semigroup property and the contribution of the lower order terms in H. Each derivative of K
ASYMPTOTICS OF SEMIGROUP KERNELS
131
gives a t−1/m -singularity. We say that Aα K has the canonical (large time) Gaussian upper bounds if the bounds (2) are valid for all t > 0 with ω = 0. On a subclass of Lie groups these bounds are valid for particular operators. 2. Second-order operators A Lie group is said to have polynomial growth if there exist c > 0 and D ∈ N0 such that the volume V 0 (ρ) ≤ c ρD for all ρ ≥ 1 ([Gui] and [Jen]). It turns out that this definition is independent of the choice of the algebraic basis. Theorem 2.1. If G has polynomial growth and H is a pure secondorder real symmetric subcoercive operator of step 2 then there exist b, c > 0 such that (2)
|Kt | ≤ c Gb,t
and
(2)
|Ai Kt | ≤ c t−1/2 Gb,t
for all t > 0 and i ∈ {1, . . . , d0 }. Proof. See [SC].
In the situation of Theorem 2.1 the reality of the coefficients implies that the kernel K is positive and by the Beurling-Deny criterium the semigroup SR is a contraction semigroup on L∞ . Therefore K is integrable and Kt = 1 for all t > 0. By duality the semigroup is also a contraction semigroup on L1 and then a Nash inequality provides L∞ -bounds on Kt . Finally by a Davies perturbation the L∞ -bounds for K can be strengthened to the Gaussian bounds of Theorem 2.1. If the coefficients are complex then K is not positive and there is no version of the Beurling-Deny criterium for complex operators [ABBO]. It is a recent result that the reality of the coefficients in Theorem 2.1 is superfluous. Theorem 2.2. If G has polynomial growth and H is a pure secondorder (complex) subcoercive operator of step 2 then there exist b, c > 0 such that (2) (2) |Kt | ≤ c Gb,t and |Ai Kt | ≤ c t−1/2 Gb,t for all t > 0 and i ∈ {1, . . . , d0 }. Proof. See [DER2].
Although the first-order derivatives of the kernel satisfy the canonical large time Gaussian upper bounds, in general higher order derivatives fail to have the canonical large time Gaussian upper bounds.
132
A.F.M. TER ELST AND DEREK W. ROBINSON
Theorem 2.3. Suppose G has polynomial growth and let H be a pure second-order (complex) subcoercive operator of step 2. The following are equivalent. I. There exist b, c > 0 such that (2)
|Ai Aj Kt | ≤ c t−1 Gb,t
for all t > 0 and i, j ∈ {1, . . . , d0 }. II. There exists a c > 0 such that kAi Aj St k2→2 ≤ c t−1 for all t > 0 and i, j ∈ {1, . . . , d0 }. III. For all p ∈ h1, ∞i there exists a c > 0 such that kAi Aj ϕkp ≤ c kHϕkp for all ϕ ∈ D(H) and i, j ∈ {1, . . . , d0 }. IV. There exist σ ∈ h0, 1i, c > 0 and for all R ∈ h0, ∞i a function ηR ∈ C ∞ (G) such that 0 ≤ ηR ≤ 1, ηR (g) = 1 for all g ∈ G with |g|0 ≤ σR, ηR (g) = 0 for all g ∈ G with |g|0 ≥ R and kAα ηR k∞ ≤ c R−|α| for all α ∈ J(d0 ) with |α| = 2. V. The Lie algebra of G is the direct product of the Lie algebra of a compact group and a nilpotent Lie algebra. Proof. For real symmetric operators this theorem has been proved in [ERS2]. The general case is again in [DER2]. In [Ale] Alexopoulos gave an example of a sublaplacian H on the covering group of the Euclidean motion group for which Condition III of Theorem 2.3 fails. A similar theorem is valid for n derivatives instead of only two derivatives, with n ≥ 3. In particular, on a nilpotent Lie group all higher order derivatives of the kernel associated to a complex second-order operator of step r, with r ≥ 2, satisfy the canonical Gaussian upper bounds for large time. That is a special case of the following theorem. Theorem 2.4. Let G be a nilpotent Lie group and let r be the rank of its Lie algebra. Let m ∈ 2N and let H be a pure m-th order operator which is subcoercive of step r. Then for all α ∈ J(d0 ) there exist b, c > 0 such that (m) |Aα Kt | ≤ c t−|α|/m Gb,t for all t > 0. Moreover, for all N ∈ N and p ∈ h1, ∞i there exists a c > 0 such that (3)
c−1 max kAα ϕkp ≤ kH N/m ϕkp ≤ c max kAα ϕkp |α|=N
for all ϕ ∈ D(H
N/m
|α|=N
)=
T
α
|α|≤N
Proof. See [NRS] or [ERS1].
D(A ).
ASYMPTOTICS OF SEMIGROUP KERNELS
133
3. Sums of subcoercive operators Before we describe a more general theorem we first consider an example on Rd . P Example 3.1. Let ∆ = − di=1 ∂i 2 be the Laplacian on Rd and n, m ∈ 2N such that n ≤ m. Set H = ∆n/2 + ∆m/2 . Let K, K (n) and K (m) denote the kernels of the semigroups generated by H, ∆n/2 and ∆m/2 . Then Kt ∗ ϕ = St ϕ = e−tH ϕ = e−t∆
n/2
e−t∆
m/2
(n)
ϕ = Kt
(m)
∗ Kt
∗ϕ (n)
(m)
for all ϕ ∈ L2 (Rd ) since ∆n/2 and ∆m/2 commute. So Kt = Kt ∗ Kt for all t > 0. Hence there exist b, c > 0 such that (n)
(m)
|Kt | ≤ c (Gb,t ∗ Gb,t ) for all t > 0 and x ∈ Rd . These bounds can be reexpressed as follows. Set (m,n)
Eb,t
(x) = (t−d/n ∧ t−d/m )(e−b(|x|
n t−1 )n/(n−1)
∨ e−b(|x|
m t−1 )m/(m−1)
) .
Then for all b > 0 there exist b0 , c > 0 such that (n)
(m)
(m,n)
Gb,t ∗ Gb,t ≤ c Eb0 ,t and (m,n)
Eb,t
(n)
(m)
≤ c (Gb0 ,t ∗ Gb0 ,t ) (m,n)
for all t > 0. Thus there are b, c > 0 such that |Kt | ≤ c Eb,t for all t > 0. Using Fourier analysis it is not hard to show that there exists a c > 0 such that (n) c−1 t−ν t−d/n ≤ kKt − Kt k∞ ≤ c t−ν t−d/n uniformly for all t ≥ 1, where ν = (m − n)/n. So for large t the kernel K (n) is a first approximation of K. One might hope that one (n) has bounds |Kt | ≤ c Gb,t for suitable b, c > 0, uniformly for all t ≥ 1, but these are not valid by the following argument. If y ∈ Rd then the Lebesgue dominated convergence theorem implies that Z m d/m 1/m −d lim t Kt (t y) = (2π) dp e−ip·y e−|p| t→∞
and the integral is not zero for all y ∈ Rd . But (n)
lim td/m Gb,t (t1/m y) = 0
t→∞
134
A.F.M. TER ELST AND DEREK W. ROBINSON
for all b > 0 and y ∈ Rd . Therefore there are no b, c > 0 such that (n) |Kt | ≤ c Gb,t uniformly for all t ≥ 1. The general case on a nilpotent Lie group is as follows. Theorem 3.2. Let G be a nilpotent Lie group and let r be the rank of its Lie algebra. Let k ∈ N\{1} and m1 , . . . , mk ∈ 2N with m1 > m2 > . . . > mk . For all j ∈ {1, . . . , k} let Hmj be a pure mj -th order P subcoercive operator of step r. Set H = kj=1 Hmj and let K be the kernel of the semigroup generated by H. Then one has the following. I. For all α there are b, c > 0 such that (m)
(m)
|Aα Kt | ≤ c (t−|α|/m ∧ t−|α|/m ) (Gb,t ∗ Gb,t ) II.
for all t > 0 where m = m1 and m = mk . For all α ∈ J(d0 ) and p ∈ h1, ∞i there exists a c > 0 such that kAα ϕkp ≤ ckH |α|/mj ϕkp
for all j ∈ {1, . . . , k} and ϕ ∈ D(H |α|/mj ). III. For all α ∈ J(d0 ) there exist b, c > 0 such that (m)
|Aα Kt − Aα Kt
(m)
(m)
| ≤ c t−ν t−|α|/m (Gb,t ∗ Gb,t )
for all t ≥ 1, where K (m) denotes the kernel of Hm and ν = (mk−1 − mk )/mk . Proof. See [DER1], Theorems 2.1 and 2.12.
Again Statement III of Theorem 3.2 indicates that K (m) is the first order approximation of the kernel K for large t. Thus the large time behaviour is determined by the lowest order terms in H. The kernel K can be bounded by a Gaussian only in a very special case. Proposition 3.3. Let n ∈ N\{1} and adopt the notation of Theorem 3.2. The following are equivalent. I. II.
(n)
There exist b, c > 0 such that |Kt | ≤ c Gb,t for all t > 0. n = m = m or G is compact and n ≥ m.
Proof. See [DER1], Proposition 2.15.
4. Higher order operators It follows from Theorem 2.4 and Example 1.1 that for all m ∈ 2N the P0 m/2 kernel of the semigroup generated by the operator − di=1 Ai 2 satisfies canonical Gaussian upper bounds for all time if G is nilpotent. The condition that G is nilpotent can be relaxed.
ASYMPTOTICS OF SEMIGROUP KERNELS
135
Theorem 4.1. Let a1 , . . . , ad0 be an algebraic basis of rank r for the Lie algebra of a Lie group G with polynomial growth. Let m ∈ 2N and let K be the kernel of the semigroup generated by the operator P0 m/2 (m) − di=1 Ai 2 . Then there exist b, c > 0 such that |Kt | ≤ c Gb,t for all t > 0. Proof. See [ElR3], Theorem 3.1.
It also follows from Theorem 2.4 and Example 1.1 that for all m ∈ 2N P0 the kernel of the semigroup generated by the operator (−1)m/2 di=1 Am i satisfies canonical Gaussian upper bounds for all time if G is nilpotent. Nevertheless in contrast to Theorem 4.1, this result does not extend to Lie groups with polynomial growth, in general. We next describe a counter example. 5. The Euclidean motion group The Lie algebra g of the Euclidean motion group is the three dimensional Lie algebra with basis b1 , b2 , b3 and commutation relations [b1 , b2 ] = b3 ,
[b1 , b3 ] = −b2 , [b2 , b3 ] = 0 .
Then g is solvable, but not nilpotent. The maximal nilpotent ideal of g, the nilradical, equals n = span(b2 , b3 ). Let G be the connected simply connected Lie group with Lie algebra g. Then G is the covering group of the Euclidean motion group. Let a1 , a2 be an algebraic basis for g, let m ∈ 2N\{2} and set (4)
m H = (−1)m/2 (Am 1 + A2 ) .
Let K be the kernel of the semigroup generated by H. One has V 0 (ρ) ρ3 for ρ ≥ 1, so G has polynomial growth. Theorem 5.1. The following are equivalent. I. There exist b, c > 0 such that (m)
|Kt | ≤ c Gb,t II.
uniformly for all t > 0. There exists a c ≥ 1 such that c−1 V (t)−1/m ≤ kKt k∞ ≤ c V (t)−1/m
uniformly for all t > 0. III. a1 ∈ n or a2 ∈ n. Proof. See [ElR3] Theorem 1.1 and Remark 2.5.
136
A.F.M. TER ELST AND DEREK W. ROBINSON
A sketch of the beginning of the proof is as follows. Define Φ : R3 → G by Φ(x1 , x2 , x3 ) = exp(x1 b1 ) exp(x2 b2 ) exp(x3 b3 ) . ˇi = (Φ−1 )∗ Bi . Then Φ is a diffeomorphism. Set Bi = dL(bi ) and B Then ˇ1 = −∂1 , B ˇ2 = − cos x1 ∂2 + sin x1 ∂3 , B ˇ3 = − sin x1 ∂2 − cos x1 ∂3 , B ˇ = (Φ−1 )∗ H. Note where the ∂i are the partial derivatives on R3. Set H that 2 X m/2 m/2 ˇ (5) (ϕ, Hϕ) = (Aˇi ϕ, Aˇi ϕ) i=1
ˇ Then the quadratic form on the right hand side of for all ϕ ∈ D(H). (5) can be written in the form X ˇ = (ϕ, Hϕ) (∂ α ϕ, cα,β ∂ β ϕ) α,β∈J(3) 1≤|α|,|β|≤m/2
with the cα,β a function which is a polynomial of functions x 7→ sin x1 ˇ might conand x 7→ cos x1 . Hence the quadratic form associated to H tain second-order terms. The large time behaviour of the kernel can be obtained using a Bloch–Zak decomposition or by using homogenization theory. Example 5.2. If H = B1 4 + B2 4 then K satisfies the canonical Gaussian upper bounds. If H = B1 4 + (B1 + B2 )4 then K does not satisfy the canonical ˇ1 + B ˇ2 )2ϕ as Gaussian upper bounds for large time. If one expands (B ˇ3 ϕ, in (5) then one obtains on both side of the inner product a term B so ˇ = (B ˇ3 ϕ, B ˇ3 ϕ) + higher order derivatives . (ϕ, Hϕ) Similarly to the situation in Theorem 3.2 for sums of subcoercive operators on nilpotent Lie groups one has a contribution of a second-order operator and K does not satisfy the m-th order canonical Gaussian upper bounds for large time. However, if H = (B1 + B2 )4 + B3 4 then K does have the canonical Gaussian upper bounds for large time, since Condition III of Theorem 5.1 is valid. This is surprising since the quadratic form has a
ASYMPTOTICS OF SEMIGROUP KERNELS
137
second-order term contribution as in the second example. But homogenization is a non-linear process and for the homogenized operator in this case the coefficients of the second-order terms cancel. It turns out ˇ is again a fourth-order operator and K that the homogenization of H satisfies the canonical fourth-order Gaussian upper bounds. Next we describe the behaviour of the kernel in case the equivalent conditions of Theorem 5.1 are not valid. Theorem 5.3. Let H be as in (4). Suppose a1 6∈ n and a2 6∈ n. Then one has the following. I. There exist b, c > 0 such that Z (m) [2] |Kt (g)| ≤ c dh Gb,t (gh−1 ) Gb,t (h) N
uniformly for all g ∈ G and t ≥ 1, where N = exp n and −1
[2]
Gb,t (Φ(0, x2 , x3 )) = t−1 e−b(x2 2+x3 2)t II.
is the second-order Gaussian on N . There exists a c > 0 such that c−1 t−(m+1)/m ≤ kKt k∞ ≤ c t−(m+1)/m
uniformly for all t ≥ 1. III. There exist c, c0 , c1 > 0 such that ˇt − K b t k∞ ≤ c t−(m+1)/m t−1/m kK ˇ is the kernel of the semigroup uniformly for all t ≥ 1, where K ˇ = (Φ−1 )∗ H and K b is the kernel of the semigroup generated by H generated by (−1)m/2 c1 ∂1m − c0 (∂2 2 + ∂3 2) . Proof. See [ElR3].
P If H = (−1)m/2 |α|=|β|=m/2 ∂ α cα,β ∂ β is a pure m-th order strongly elliptic operator on Rd with complex measurable coefficients then the solution of the Kato problem solved by Auscher, Hofmann, McIntosh and Tchamitchian states that D(H 1/2 ) = W m/2,2 and there exists a c > 0 such that (6)
c−1 max k∂ α ϕk2 ≤ kH 1/2 ϕk2 ≤ c max k∂ α ϕk2 |α|=m/2 1/2
|α|=m/2
for all ϕ ∈ D(H ) (see [AHMT], Theorem 1.5). In [AHMT] the identity D(H 1/2 ) = W m/2,2 and the homogeneous estimates (6) were proved first under the additional assumption that the kernel of the semigroup generated by H has the canonical Gaussian upper bounds.
138
A.F.M. TER ELST AND DEREK W. ROBINSON
A similar result is valid for subcoercive operators with constant coefficients on nilpotent groups by (3) in Theorem 2.4. If the group is simply connected then an operator as in Theorem 2.4 is unitarily equivalent to an operator on Rd with polynomial coefficients. If G is a general Lie group and H is a pure m-th order operator which generates a bounded semigroup on L2 then it follows from Theorem 1.2.V that there exists a c > 0 such that c−1 ( max kAα ϕk2 + kϕk2 ) ≤ kH 1/2 ϕk2 + kϕk2 |α|=m/2
≤ c ( max kAα ϕk2 + kϕk2 )
(7)
|α|=m/2
for all ϕ ∈ D(H 1/2 ). But if the group is not homogeneous then one cannot easily scale the L2 -norm kϕk2 of ϕ away in (7). Nevertheless the homogeneous estimates (3) are valid on nilpotent Lie groups, even if the nilpotent group is not homogeneous. If G is the covering group of the Euclidean motion group and H = B1 4 + B2 4, with the notation as in the beginning of this section, then it follows from Theorem 5.1 that the kernel of the semigroup S generated by H has the canonical Gaussian upper bounds. Moreover, S is T bounded on L2 and D(H 1/2 ) = |α|=2 D(Aα ) on L2 . Nevertheless the analogue of the homogeneous estimates (6) and (3) are not valid. Proposition 5.4. Let G be the covering group of the Euclidean motion group and adopt the notation as in the beginning of this section. Let a1 , a2 be an algebraic basis for g such that a1 ∈ n or a2 ∈ n. Let m m ∈ 2N\{2} and set H = (−1)m/2 (Am 1 + A2 ). Then there does not exist a c > 0 such that max kAα ϕk2 ≤ c kH 1/2 ϕk2
|α|=m/2
for all ϕ ∈ D(H 1/2 ). Proof. Suppose there exists a c > 0 such that max|α|=m/2 kAα ϕk2 ≤ c kH 1/2 ϕk2 for all ϕ ∈ D(H 1/2 ). We will show that Condition IV of Theorem 2.3 is valid. Let K be the kernel of the semigroup S generated by H. Then max kAα St ϕk2 ≤ c kH 1/2 St ϕk2 ≤ (2e)−1/2 t−1/2 kϕk2
|α|=m/2
for all t > 0 by spectral theory. By Theorem 5.1 the kernel K satisfies the canonical Gaussian upper bounds. Hence by a quadrature estimate it follows that there exists a c1 > 0 such that kSt k1→2 ≤ c1 V 0 (t)−1/(2m)
and
kKt k2 ≤ c1 V 0 (t)−1/(2m)
ASYMPTOTICS OF SEMIGROUP KERNELS
139
for all t > 0. Therefore one has kAα Kt k∞ ≤ kAα S3t k1→∞ ≤ kAα S2t k2→∞ kSt k1→2 = kAα K2t k2 kSt k1→2 ≤ kAα St k2→2 kKt k2 kSt k1→2 ≤ (2e)−1/2 c c1 2 t−1/2 V 0 (t)−1/m for all t > 0 and α ∈ J(d0 ) with |α| = m/2. Hence there exists a c2 > 0 such that max kAα Kt k∞ ≤ c2 t−1/2 V 0 (t)−1/m
(8)
|α|=m/2
for all t > 0. Since H does not have a constant term one has H11 = 0 on L∞ , where 11 is the constant function with value one. Hence St 11 = 11 and R Kt = 1 for all t > 0. Moreover, since H is self-adjoint one deduces that Kt (g −1 ) = Kt (g) for all g ∈ G and t > 0. Hence Z Re K2t (g) = Re dh Kt (h) Kt (h−1 g) Z ≤
Z
2
dh |Kt (h)| =
dhKt (h) Kt (h−1 ) = K2t (e)
for all t > 0 and g ∈ G by the Schwartz inequality. By Theorem 5.1 (m) there exist b, c3 > 0 such that |Kt | ≤ c3 Gb,t for all t > 0. Then for all κ > 0 one has Z 0 1/m −1 Kt (e) ≥ V (κt ) dg Re Kt (g) {g∈G:|g|0 ≤κt1/m } 0
1/m −1
= V (κt
)
Z 1−
dg Re Kt (g)
{g∈G:|g|0 >κt1/m } 0
1/m −1
≥ V (κt
)
Z 1− {g∈G:|g|0 >κt1/m }
(m) dg c3 Gb,t (g)
for all t > 0. But the last integral tends to zero as κ → ∞. Hence there exists a κ > 0 such that Kt (e) ≥ 2−1 V 0 (κt1/m )−1 for all t > 0. But since G has polynomial growth there then exists a c4 > 0 such that Kt (e) ≥ c4 V 0 (t)−1/m for all t > 0. It follows from a subelliptic variation of Lemma III.3.3 of [Rob] that there exists a c5 > 0 such that max kAα ϕk∞ ≤ εm/2−n max kAα ϕk∞ + c5 ε−n kϕk∞
|α|=n
|α|=m/2
T for all n ∈ {1, . . . , m/2 − 1} and ϕ ∈ |α|=m/2 D(Aα ) in the L∞ -sense. Hence by (8), using the Gaussian upper bounds for K and choosing
140
A.F.M. TER ELST AND DEREK W. ROBINSON
ε = t−n/m one deduces that max kAα Kt k∞ ≤ (c2 + c3 c5 ) t−n/m V 0 (t)−1/m
|α|=n
for all t > 0 and n ∈ {1, . . . , m/2}. Then 0
|Kt (g) − Kt (e)| ≤ |g|
d0 X
1/2 kAi Kt k∞ 2 ≤ c6 |g|0 t−1/m V 0 (t)−1/m
i=1
for all g ∈ G and t > 0, where c6 = (d0 )1/2 (c2 + c3 c5 ). It follows that Kt (g) −1 −b((|g|0 )m t−1 )1/(m−1) ≥ 1 − |Kt (g) − Kt (e)| c3 c4 e ≥ Kt (e) Kt (e) 0 −1/m ≥ 1 − c6 c−1 4 |g| t
(9)
for all g ∈ G and t > 0. Next let τ : C → R be a C ∞ -function such −b that 0 ≤ τ ≤ 1, τ (z) = 0 for all |z| ≥ c3 c−1 and τ (z) = 1 for all 4 e −1 −b −1 ∞ |z| ≤ 2 c3 c4 e . For all R > 0 define ηR ∈ C (G) by ηR (g) = τ
K
Rm (g)
KRm (e)
.
Then it follows from (9) that ηR (g) = 0 if |g|0 ≥ R and ηR (g) = 1 if −1 −b −1 |g|0 ≤ σR where σ = c4 c−1 6 (1 − 2 c3 c4 e ). Next we show that the derivatives have the right decay. Let α ∈ J(d0 ) with |α| = m/2. Then (10)
α
(A ηR )(g) =
X
τ
(l)
KRm (g) KRm (e)
Y l
(Aβp KRm )(g) KRm (e) p=1
uniformly for all g ∈ G and R > 0, where the sum is finite and over a subset of all l ∈ {1, . . . , n} and β1 , . . . , βl ∈ J(d0 ) with |βp | ≥ 1 for all p ∈ {1, . . . , l} and |β1 | + . . . + |βl | = n. Then l l Y (Aβp KR2 )(g) Y −|βp | −n ≤ (c2 + c3 c5 )c−1 = (c2 + c3 c5 )l c−l 4 R 4 R KR2 (e) p=1
p=1
uniformly for g ∈ G and R > 0. Hence Condition IV of Theorem 2.3 is valid. Therefore by Theorem 2.3 g is the direct product of the Lie algebra of a compact group and a nilpotent Lie algebra. This is a contradiction and the proof of the proposition is complete.
ASYMPTOTICS OF SEMIGROUP KERNELS
141
Acknowledgements This paper has been written whilst the first named author was on sabbatical leave visiting the Centre for Mathematics and its Applications at the ANU. He wishes to thank the CMA for its hospitality and the support from a grant of the Australian Research Council. References ´lemy, L., Be ´nilan, P. and Ouhabaz, E.M., [ABBO] Auscher, P., Barthe Absence de la L∞ -contractivit´e pour les semi-groupes associ´es aux op´erateurs elliptiques complexes sous forme divergence. Potential Anal. 12 (2000), 160–189. [AHMT] Auscher, P., Hofmann, S., McIntosh, A. and Tchamitchian, P., The Kato square root problem for higher order elliptic operators and systems on Rn . J. Evol. Equ. 1 (2001), 361–385. [Ale] Alexopoulos, G., An application of homogenization theory to harmonic analysis on solvable Lie groups of polynomial growth. Pacific J. Math. 159 (1993), 19–45. [BER] Burns, R.J., Elst, A.F.M. ter and Robinson, D.W., Lp -regularity of subelliptic operators on Lie groups. J. Operator Theory 31 (1994), 165–187. [DER1] Dungey, N., Elst, A.F.M. ter and Robinson, D.W., Asymptotics of sums of subcoercive operators. Colloq. Math. 82 (1999), 231–260. [DER2] , Analysis on Lie groups with polynomial growth, 2002. In preparation. [ElR1] Elst, A.F.M. ter and Robinson, D.W., Subcoercivity and subelliptic operators on Lie groups II: The general case. Potential Anal. 4 (1995), 205–243. , Weighted subcoercive operators on Lie groups. J. Funct. Anal. 157 [ElR2] (1998), 88–163. , On anomalous asymptotics of heat kernels. In Lumer, G. and [ElR3] Weis, L., eds., Evolution equations and their applications in physical and life sciences, vol. 215 of Lecture Notes in Pure and Applied Mathematics. Marcel Dekker, New York, 2001, 89–103. [ERS1] Elst, A.F.M. ter, Robinson, D.W. and Sikora, A., Heat kernels and Riesz transforms on nilpotent Lie groups. Coll. Math. 74 (1997), 191–218. [ERS2] , Riesz transforms and Lie groups of polynomial growth. J. Funct. Anal. 162 (1999), 14–51. [Gui] Guivarc’h, Y., Croissance polynomiale et p´eriodes des fonctions harmonique. Bull. Soc. Math. France 101 (1973), 333–379. [Jen] Jenkins, J.W., Growth of connected locally compact groups. J. Funct. Anal. 12 (1973), 113–127. [NRS] Nagel, A., Ricci, F. and Stein, E.M., Harmonic analysis and fundamental solutions on nilpotent Lie groups. In Sadosky, C., ed., Analysis and partial differential equations, Lecture Notes in pure and applied Mathematics 122. Marcel Dekker, New York etc., 1990, 249–275.
142
[Rob] [SC] [VSC]
A.F.M. TER ELST AND DEREK W. ROBINSON
Robinson, D.W., Elliptic operators and Lie groups. Oxford Mathematical Monographs. Oxford University Press, Oxford etc., 1991. Saloff-Coste, L., Analyse sur les groupes de Lie `a croissance polynˆ omiale. Arkiv f¨ or Mat. 28 (1990), 315–331. Varopoulos, N.T., Saloff-Coste, L. and Coulhon, T., Analysis and geometry on groups. Cambridge Tracts in Mathematics 100. Cambridge University Press, Cambridge, 1992.
Department of Mathematics and Computing Science, Eindhoven University of Technology, P.O. Box 513, 5600 MB Eindhoven, The Netherlands E-mail address: [email protected] Centre for Mathematics and its Applications, Mathematical Sciences Insitute, Australian National University, Canberra, ACT 0200, Australia E-mail address: [email protected]
THE T (b) THEOREM AND ITS VARIANTS T. TAO Abstract. We survey the various local and global variants of T (1) and T (b) theorems which have appeared in their literature, and outline their proofs based on a local wavelet coefficient perspective. This survey is based on recent work in [4] with Pascal Auscher, Steve Hofmann, Camil Muscalu, and Christoph Thiele.
1. Introduction The purpose of this expository article is to survey the various types of T (1) and T (b) theorems which have been used to prove boundedness of Calder´on-Zygmund kernels. These theorems have many applications to Cauchy integrals and analytic capacity, and also to elliptic and accretive systems; see e.g. [1], [6], [9], [13]. However we will not concern ourselves with applications here, and focus instead on the mechanics of proof of these theorems. In particular we present a slightly nonstandard method of proof, based on pointwise estimates of wavelet coefficients. For simplicity we shall work just on the real line IR, with the standard Lebesgue measure dx. For applications one must consider much more complicated settings, for instance one-dimensional sets endowed with Hausdorff measure, but we will not discuss these important generalizations in this article1, and instead concentrate our attention on the distinction between T (1) and T (b) theorems, and between local and global variants of these theorems. To avoid technicalities (in particular, in justifying whether integrals actually converge) we shall restrict all functions and operators to be real-valued, and only consider Calder´on-Zygmund operators T of the form Z T f (x) := K(x, y)f (y) dy
1In
particular we will not discuss the important recent extensions of the T (1) and T (b) theory to non-doubling measures, see e.g. [24], [23]. 143
144
T. TAO
where K(x, y) is a smooth, compactly supported kernel obeying the estimates |K(x, y)| . 1/|x − y| |K(x, y) − K(x0 , y)| + |K(y, x) − K(y, x0 )| . |x − x0 |/|x − y|2
(1) (2)
for all distinct x, x0 , y. Here we use A . B or A = O(B) to denote the estimate A ≤ CB for some constant C. An example such an R of 1 operator would be the Hilbert transform Hf (x) := p.v. x−y f (y) dy, 1 to be smooth and compactly after first truncating the kernel p.v. x−y supported. Of course, since we have declared K to be smooth and compactly supported, the operator T is a priori bounded on Lp (IR) for 1 < p < ∞. However, the basic problem is to find useful conditions under which one can obtain more quantitative bounds on the Lp operator norm of T (e.g. depending only on p and the implicit constants in (1), (2) and some additional data, but not on the smoothness and compact support assumptions on K). The remarkable T (1) theorem of David and Journ´e [12] gives a complete characterization of this problem: Theorem 1.1 (Global T (1) theorem). [12] If T is a Calder´ on-Zygmund operator such that • (Weak boundedness property) One has hT χI , χI i = O(|I|) for all intervals I. Here |I| denotes the length of I. • We have kT (1)kBM O . 1. • We have kT ∗ (1)kBM O . 1. Then T is bounded on Lp , 1 < p < ∞ kT f kp . kf kp (the implicit constants depend on p) and we have the L∞ to BM O bounds kT f kBM O . kf k∞ kT ∗ f kBM O . kf k∞ . The remarkable thing about this theorem is that while the conclusion asserts that T and T ∗ obey certain bounds for all Lp functions f (including when p = ∞), whereas the hypotheses only require these type of bounds for very specific Lp functions f . The point is that the kernel bounds (1), (2) impose severe constraints on the behavior of T . Informally, these bounds assert that the bilinear form hT f, gi is small when f and g have widely separated supports, and is especially small when one also assumes that f or g has mean zero. This allows us
THE T (b) THEOREM AND ITS VARIANTS
145
in many cases to estimate hT f, gi by replacing f with its mean on a certain interval (i.e. by replacing f with its projection onto constant functions), or by performing a similar replacement for g. By doing this at all scales simultaneously (by use of such tools as the wavelet decomposition), we can eventually control hT f, gi by quantities involving T (1), T ∗ (1), or quantities such as hT χI , χI i. In more precise terms, the standard proof of Theorem 1.1 proceeds by decomposing T into an easily estimated “diagonal” component (controlled by hT χI , χI i, and two paraproducts, one related to T (1) and one related to T ∗ (1), which can then be estimated with the aid of the Carleson embedding theorem. In this note we give an alternate proof (in a model case) based on pointwise estimates of wavelet coefficients. Despite giving a complete answer to our problem, Theorem 1.1 is not completely satisfactory for two reasons. Firstly, the conditions kT (1)kBM O . 1 and kT ∗ (1)kBM O . 1 are global rather than local, in that the constant function 1 and the nature of the BMO norm are both spread out over all intervals of space simultaneously, in a way that the weak boundedness property is not. Thus if one were only interested in whether T is bounded on acertain subset Ω of IR, the hypotheses of this theorem would be inappropriate. Secondly, the theorem is rather inflexible in that one must test T and T ∗ against the function 1, rather than some other functions which might be more convenient to apply T and T ∗ to. The first objection is easily addressed. Indeed, it is quite straightforward (and well-known) to see that the global T (1) theorem is equivalent to the following local version: Theorem 1.2 (Local T (1) theorem). If T is a Calder´ on-Zygmund operator such that R • RI |T χI (x)|2 dx . |I| for all intervals I. • I |T ∗ χI (x)|2 dx . |I| for all intervals I. Then the conclusions of Theorem 1.1 hold. We sketch the derivation of the global T (1) theorem from the local T (1) theorem as follows. Suppose that T (1) ∈ BM O, then for any interval I we would have kT (1) − [T (1)]I kL2 (I) . |I|1/2 , R 1 where [f ]I := |I| f denotes the mean of f on I. On the other hand, I from (1), (2) and a direct calculation one can show that kT (1 − χI ) − [T (1 − χI )]I kL2 (I) . |I|1/2 ,
146
T. TAO
and hence that kT (χI ) − [T (χI )]I kL2 (I) . |I|1/2 . On the other hand, from the weak boundedness property we have [T (χI )]I = O(1), and thus we have verified the first property of Theorem 1.2. The second property is verified similarly and we are done. The reverse implication (that the global theorem implies the local version) is similar and we do not detail it here. As the name suggests, the local T (1) theorem can be localized. Here is one sample localization result: Theorem 1.3 (Localized T (1) theorem). Let Ω ⊂ IR be an open set. If T is a Calder´on-Zygmund operator such that R • RI |T χI (x)|2 dx . |I| for all intervals I such that I ∩ Ω 6= ∅. • I |T ∗ χI (x)|2 dx . |I| for all intervals I such that I ∩ Ω 6= ∅. Then T is bounded on Lp (Ω) for 1 < p < ∞, in the sense that kT f |Ω kLp (Ω) . kf kLp (Ω) for all f ∈ Lp (Ω). Corresponding Lp and L∞ → BM O estimates can also be made, but are a little technical due to the need to define BMO on domains. We prove this theorem in a model case in Section 3. Unfortunately, this localization is imperfect, since the intervals I in the hypotheses are allowed to partially extend outside Ω; we do not know how to resolve this issue to make the localization more satisfactory. Now we discuss the second objection to the T (1) theorem, that the choice of function 1 is fixed. Certainly we cannot replace 1 by an arbitrary L∞ function, since if one replaces 1 by (for instance) the zero function 0 then the hypotheses become trivial, and the theorem absurd. Or if one replaces 1 by a function which vanishes on a large set, then the hypothesis yields us very little information about T on that set, and so one would not expect to conclude boundedness on T inside that set. However, if one replaces 1 by functions which in some sense “never vanish”, then one can generalize the T (1) theorem: Theorem 1.4 (Global T (b) theorem). [13] Let b, b0 be bounded in L∞ (IR) and such that we have the pseudo-accretivity condition |[b]I |, |[b0 ]I | & 1 for all intervals I. If T is a Calder´ on-Zygmund operator such that • (Modified weak boundedness property) One has hT (bχI ), (b0 χI )i = O(|I|) for all intervals I.
THE T (b) THEOREM AND ITS VARIANTS
147
• We have kT (b)kBM O . 1. • We have kT ∗ (b0 )kBM O . 1. Then T is bounded on L2 kT f k2 . kf k2 and we have the L∞ to BM O bounds kT f kBM O . kf k∞ kT ∗ f kBM O . kf k∞ . The conditions b, b0 be bounded in L∞ (IR) can be relaxed slightly to requiring b, b0 to just be bounded in BMO; see [4]. This theorem can be proven by using wavelet systems adapted to b, b0 respectively [5]. In the special “one-sided” case when b0 = 1 then there is an alternate proof going through the global T (1) theorem and the heuristic that when T ∗ (1) ∈ BM O, we have the approximation hT (b), φI i ≈ [b]I hT (1), φI i for all intervals I and all bump functions φI of mean zero adapted to I; see [25] (and also [4]). This T (b) theorem can also be localized, so that instead of having a single function b supported globally, we only need a localized function bI for each I. Also, each bI only has to satisfy a single accretivity condition: Theorem 1.5 (Local T (b) theorem). [4] If T is a Calder´ on-Zygmund operator such that • For every interval I, there exists a function bI ∈ L2 (I) with Z |bI |2 + |T bI |2 . |I| I
which obeys the accretivity condition |[bI ]I | & 1. • For every interval I, there exists a function b0I ∈ L2 (I) with Z |b0I |2 + |T ∗ b0I |2 . |I| I
which obeys the accretivity condition |[b0I ]I | & 1. Then T is bounded on L2 kT f k2 . kf k2 and we have the L∞ to BM O bounds kT f kBM O . kf k∞
148
T. TAO
kT ∗ f kBM O . kf k∞ . Strictly speaking, the theorem in [4] was only proven in a model case in which “perfect” kernel cancellation conditions were assumed, but the argument extends without significant difficulty to general Calder´onZygmund kernels. This theorem is a variant of an earlier local T (b) theorem of [9], which was in a much more general setting but required L∞ conditions on bI , T bI , b0I , T ∗ b0I rather than L2 conditions. One can 0 also replace the L2 conditions with Lp and Lp conditions, see [4]. In the “one-sided case”, when T ∗ 1 ∈ BM O, then one does not need the second hypothesis (involving the b0I ), and the proof can again proceed via the T (1) theorem, again using the heuristic that when T ∗ 1 ∈ BM O, we have the approximation hT bI , φI i ≈ [bI ]I hT (1), φI i when φI has mean zero and is adapted to I. (cf. the local T (b) theorems in [6], [1], [2]). However in the general case it appears that one is forced to use tools such as adapted wavelet systems. 2. A dyadic model In order to simplify the exposition, we shall replace our “continuous” Calder´on-Zygmund operators with a “discrete” dyadic model; this model will be much cleaner to manipulate because of the absence of several minor error terms, but will already capture the essence of the arguments. We define a dyadic interval to be any interval of the form [2j k, 2j (k + 1)] where j, k are integers. We will always ignore sets of measure zero, and so will not distinguish between open, closed, or half-open dyadic intervals. If I is an interval, we use |I| to denote its length and I lef t and I right to denote left and right halves of I respectively, and refer to I lef t and I right as siblings. We define the Haar wavelet φI to be the L2 -normalized function φI := |I|−1/2 (χI lef t − χI right ). As is well-known, these wavelets form an orthonormal basis of L2 (IR). For any locally integrable f , we define the wavelet transform W f , defined on the space of dyadic intervals, by W f (I) := hf, φI i. Thus W is an isometry from L2 to l2 . Observe that if T obeys (1), (2), and I and J are disjoint dyadic intervals, then the quantity hT φI , φJ i decays quickly in the separation of I and J (for instance, we have the estimates −2 |I| dist(I, J) hT φI , φJ i . 1+ |J| |J|
THE T (b) THEOREM AND ITS VARIANTS
149
when |J| ≥ |I|). We now introduce a simplified model operator in which the quantity hT φI , φJ i not only decays, but in fact vanishes when I, J are disjoint. Definition 2.1. A perfect dyadic Calder´ on-Zygmund operator T is an operator of the form Z T f (x) := K(x, y)f (y) dy where K(x, y) is a bounded, compactly supported function which obeys the kernel condition 1 |K(x, y)| . (3) |x − y| and the perfect dyadic Calder´ on-Zygmund conditions |K(x, y) − K(x0 , y)| + |K(y, x) − K(y, x0 )| = 0
(4)
whenever x, x0 ∈ I and y ∈ J for some disjoint dyadic intervals I and J. Equivalently, K is constant on all rectangles {I × J : I, J are siblings}. As a consequence of (4) we see that T fI is supported on I whenever fI is supported on I with mean zero, and similarly with T replaced by T ∗ . In particular we have hT φI , φJ i = 0 whenever I, J are disjoint. We define the dyadic BMO norm by 1/2 X 1/2 Z 1 2 2 = sup |f − [f ]I | kf kBM O∆ := sup W f (J) |I| I I I J⊆I where I ranges over all dyadic intervals. This norm is very close to the usual BMO norm, see [18]. 3. Pointwise wavelet coefficient estimates The purpose of this section is to give a somewhat non-standard proof of the T (1) theorem in the model case of perfect dyadic Calder´onZygmund operators. This differs from the usual proof (involving paraproduct decomposition and Carleson embedding) in that it relies on pointwise estimates for wavelet coefficients, and also tackles the endpoint L∞ → BM O estimates first (instead of the usual procedure of first establishing L2 bounds. Let T be a perfect dyadic Calder´on-Zygmund operator, and let f ∈ 2 L (IR) be a function. In order to study the L2 or BMO norm of T f , it makes sense to study the wavelet coefficients W (T f )(J). It turns out that one can write the coefficients of T f rather explicitly in terms of the corresponding coefficients of f , T χJ , T ∗ χJ and f T ∗ χJ :
150
T. TAO
Lemma 3.1. [4] We have W (T f )(J) = [f ]J W (T χJ )(J) − [f ]J W (T ∗ χJ )(J) + W (f T ∗ χJ )(J) 2 − (hT χJlef t , χJright i + hT χJright , χJlef t i)W f (J). (5) |J| From (3) we thus have the useful pointwise estimate |W (T f )(J)| . |[f ]J ||W (T χJ )(J)| + |[f ]J ||W (T ∗ χJ )(J)| + |W (f T ∗ χJ )(J)| + |W f (J)|
(6)
on the wavelet coefficient of T f . The pointwise estimates (6) are reminiscent of the standard pointwise sharp function estimates (T f )# (x) . M f (x) for Calder´on-Zygmund kernels bounded on Lp ; see [26]. Proof First observe that if f is vanishes on J then both sides of (5) vanish (since W (T f )(J) = hf, T ∗ φJ i and T ∗ φJ is supported on J). Thus by linearity we may assume that f is supported on J. Now suppose that f is equal to χJ , then W f (J) = 0, and (5) collapses to W (T χJ )(J) = W (T χJ )(J) − W (T ∗ χJ )(J) + W (χJ T ∗ χJ )(J) which is clearly true. By linearity we may thus assume that f is orthogonal to χJ . Now suppose that f is equal to φJ . Then (5) simplifies to 1 2 hT φJ , φJ i = hT χJ , χJ i − (hT χJlef t , χJright i + hT χJright , χJlef t i), |J| |J| which is easily verified by expanding out T . By linearity we may thus assume that f is also orthogonal to φJ . In particular, we can now assume that f has mean zero on both Jlef t and Jright . After some re-arranging, the claim now collapses to hT f, φJ i = hT (f φJ ), χJ i. If f is supported on Jlef t , then the claim follows since φJ is constant on the support of f and on T f . Similarly if f is supported on Jright . The claim now follows by linearity. Of course, this identity took full advantage of the perfect cancellation (4). However in the continuous case there are still analogues of (6), but with some additional error terms which are easily treated. (Basically, the right-hand side will not just contain terms related to J, but also terms related to intervals J 0 which are “close” in size and location to J, but with an additional weight which decays fairly quickly as J 0 moves further away from J). One can also replace the Haar wavelet with smoother wavelets in order to avoid certain (harmless) logarithmic
THE T (b) THEOREM AND ITS VARIANTS
151
divergences which can come up in this procedure. We will not pursue the details. With (6) in hand it is a fairly easy matter to prove the dyadic analogue of Theorem 1.2: Theorem 3.2 (Local dyadic T (1) theorem). If T is a perfect dyadic Calder´on-Zygmund operator such that R • RI |T χI (x)|2 dx . |I| for all dyadic intervals I. • I |T ∗ χI (x)|2 dx . |I| for all dyadic intervals I. Then T is bounded on L2 kT f k2 . kf k2 and we have the L∞ to dyadic BM O bounds kT f kBM O∆ . kf k∞ kT ∗ f kBM O∆ . kf k∞ . Of course, this local theorem is equivalent to its global counterpart by the argument sketched in the introduction. Proof First we show that T maps L∞ into dyadic BM O. Fix f ∈ L∞ ; we may normalize kf k∞ = 1. We need to show the Carleson measure condition X |W (T f )(J)|2 . |J| (7) J⊆I
for all dyadic intervals I. By a standard iteration procedure (extremely common in the study of Carleson measures or BMO; we will use it again in the next section) we may restrict the summation to those intervals J for which |[T ∗ χI ]J | ≤ C
(8)
for some large constant C. We sketch for this is as follows. R the∗ reason 2 By Cauchy-Schwartz, we see that J |TR χI | ≥ C 2 |J| whenever (8) fails. Since we are assuming the bound I |T ∗ χI |2 . |I|, the intervals J for which (8) fails can therefore only cover a set of measure at most 1 |I|, if C is chosen sufficiently large. One can then iterate away those 100 contributions in the usual manner (e.g. by assuming inductively that (7) already holds for all sub-intervals of I). Fix I; we will implicitly assume that (8) holds. Now we use (6) to decompose the wavelet coefficients W (T f )(J). We write W (f T ∗ χJ )(J) = W (f T ∗ χI )(J) − W (f T ∗ χI\J )(J).
152
T. TAO
From (4) we see that T ∗ χI\J is constant on J, thus W (f T ∗ χI\J )(J) = ([T ∗ χI ]J − [T ∗ χJ ]J )W f (J). From the hypotheses and Cauchy-Schwarz we have [T ∗ χJ ]J = O(1). From this and (8) we thus have |W (f T ∗ χJ )(J)| . |W (f T ∗ χI )(J)| + |W f (J)|. Also, since T ∗ φJ is supported on J, we have W (T χJ )(J) = W (T χI )(J), and similarly for T ∗ . Since [f ]J = O(kf k∞ ) = O(1), we thus see from (6) that we have the pointwise estimate |W (T f )(J)| . |W (T χI )(J)| + |W (T ∗ χI )(J)| + |W (f T ∗ χI )(J)| + |W (f )(J)|. By hypothesis, the functions T χI , T ∗ χI , (T ∗ χI )f , and f all have an L2 (I) norm of O(|I|1/2 ). The claim (7) follows since the φJ are orthonormal in L2 (I). This shows that T maps L∞ into BM O∆ . A similar argument shows that T ∗ maps L∞ to BM O∆ . One then concludes Lp boundedness by Fefferman-Stein interpolation (see e.g. [26]) and a standard reiteration argument (starting, for instance, with the apriori L2 boundedness and then improving this by repeated interpolation; cf. [29]). Alternatively, one could use Lp Calder´on-Zygmund techniques as in [16] to first establish weak Lp bounds and then interpolate. We omit the details We now sketch how the above theorem localizes to a domain Ω ⊂ IR. Specifically, we prove Corollary 3.3 (Dyadic localized T (1) theorem). Let Ω ⊂ IR be an open set. Let I be the set of dyadic intervals I such that I ∩ Ω 6= ∅. If T is a Calder´on-Zygmund operator such that R • RI |T χI (x)|2 dx . |I| for all I ∈ I. • I |T ∗ χI (x)|2 dx . |I| for all I ∈ I. Then T is bounded on Lp (Ω) for 1 < p < ∞, in the sense that kT f |Ω kLp (Ω) . kf kLp (Ω) for all f ∈ Lp (Ω). Proof Let Π be the projection Πf :=
X I∈I
W f (I)φI .
THE T (b) THEOREM AND ITS VARIANTS
153
Equivalently, we have Πf (x) = f (x) if x ∈ Ω, and Πf (x) = [f ]I when x 6∈ Ω, where I is the smallest dyadic interval containing x and intersecting Ω. We consider the operator ΠT Π. It is easy to verify that this operator is also a perfect dyadic Calder´on-Zygmund operator. We claim that ΠT Π verifies the hypotheses of Theorem 3.3. The claim then follows since ΠT Πf (x) = T Πf (x) = T f (x) for all x ∈ Ω and f ∈ Lp (Ω). Let I be a dyadic interval. We shall verify that Z |ΠT ΠχI (x)|2 dx . |I|. I
First suppose that I ∈ I. Then ΠχI = χI . Also, on I one can estimate Π by the Hardy-Littlewood maximal function on I. The claim then follows from the hypotheses and the Hardy-Littlewood maximal inequality. Now suppose that I 6∈ I. Let J be the smallest dyadic interval in I |I| which contains J. Then ΠχI = |J| χJ , and Πf = [f ]J on I. Thus it will suffice to show that Z 2 1 |I| T χ J . |I|. |J| J |J| But this follows from hypothesis since |I| ≤ |J|. It is also possible to establish the above corollary directly by using variants of the pointwise estimate (6), combined with the squarefunction wavelet characterization of Lp and some Calder´on-Zygmund theory, but we will not detail this here. The operator Π is an example of a phase space projection, to a region of phase space known as a “tree”. Such localized estimates for Calder´on-Zygmund operators are very useful in the study of multilinear objects such as the bilinear Hilbert transform and the Carleson maximal operator: see e.g. [15], [28], [19], [20], [22], etc. (For a further discussion of the connection between Carleson measures, Calder´onZygmund theory, and phase space analysis, see [4]). 3.1. Adapted Haar bases, and T (b) theorems. We now turn to T (b) theorems, starting with the global T (b) theorem in Theorem 1.4. We shall consider the dyadic model case when T is a perfect dyadic Calder´on-Zygmund operator, and assume that b, b0 are bounded and obey the pseudo-accretivity conditions |[b]I |, |[b0 ]I | & 1 for all dyadic intervals I.
154
T. TAO
Ideally, one would like to have an estimate like (6), but with T (χJ ) and T ∗ (χJ ) replaced by T (bχJ ) and T ∗ (b0 χJ ): |W (T f )(J)| . |[f ]J ||W (T (bχJ ))(J)| + |[f ]J ||W (T ∗ (b0 χJ ))(J)| + |W (f T ∗ (b0 χJ ))(J)| + |W f (J)|.
(9)
This is because one can then approximate T (bχJ ) and T ∗ (b0 χJ ) by T (b) and T ∗ (b0 ) by arguments similar to that used to prove Theorem 3.2. Unfortunately, such an estimate does not seem to hold in general. However, if one replaces the Haar wavelet system by modified Haar wavelet systems adapted to b and b0 , then one can recover a formula similar to (9), as follows. For each interval I, we define the adapted Haar wavelet φbI (introduced in [10]; see also [5]) by φbI := φI −
W b(I) χI . [b]I |I|
(10)
The wavelet φbI no longer has mean zero, but still obeys the weighted mean zero condition Z bφbI = 0. (11) As a consequence we see that Z φbI bφbJ = 0 for all I 6= J.
(12)
A calculation gives the identity Z 2 φbI bφbI = −1 . [b]Ilef t + [b]−1 Iright From the pseudo-accretivity and boundedness properties, we thus see that φbP has the non-degeneracy property Z b b φ bφ (13) P P ∼ 1. Define the dual adapted Haar wavelet ψIb by ψIb := R
φbI b . φbI bφbI
By (12), (13) we thus have that hψIb , φbJ i = δIJ where δ is the Kronecker delta. In particular we have the representation formula X f= Wb f (I)ψIb (14) I
THE T (b) THEOREM AND ITS VARIANTS
155
(formally, at least), where the adapted wavelet coefficients Wb f (I) are defined by Wb f (I) := hf, φbI i. From some computation and the Carleson embedding theorem one can eventually verify the orthogonality property 1/2 X 2 |Wb f (I)| ∼ kf k2 (15) I
(see e.g. [26], or [4]). Thus the adapted wavelet transform Wb has most of the important properties that W has. The analogue of (9) is Lemma 3.4. [4] If f is bounded in L∞ , then |Wb (bT ∗ f )(I)| . |Wb (bT ∗ (b0 χI ))(I)| + |Wb (b0 T (bχI ))(I)| + |Wb (f T (bχI ))(I)| + |W f (I)| + |W (b0 )(I)|. Proof We can write W b(bT ∗ f )(I) = hf, T (bφbI )i = hf, T ψIb i. Since ψIb has mean zero, T ψIb is supported on I. Thus the left-hand side does not depend on the values of f outside I. Similarly for the right-hand side. Thus we may assume that f is supported on I. The claim is clearly true when f is equal to a bounded multiple of b0 χI (since the left-hand side is then bounded by the first term on the right-hand side). Since |[b0 ]I | ∼ 1, every bounded function f on I can be split as the sum of a bounded multiple of b0 χI and a bounded function on I with mean zero. Thus it will suffice to prove the estimate |Wb (bT ∗ f )(I)| . |Wb (f T (bχI ))(I)| + |W f (I)| for functions f of mean zero on I. We can write Wb (bT ∗ f )(I) = hφbI T ∗ (f ), bχI i. Since hT ∗ (φbI f ), bχI i = hf T (bχI ), φbI i = Wb (f T (bχI ))(I) it will suffice to prove the commutator estimate hφbI T ∗ (f ), bχI i − hT ∗ (φbI f ), bχI i = O(|W f (I)|). Recall that φbI is constant on Ilef t and Iright . Thus if f is supported on Ilef t with mean zero, then the commutator vanishes (since T ∗ (φbI f ) = φbI (Ilef t )T ∗ (f ) is supported on Ilef t ). Similarly if f is supported on Iright with mean zero. Since we are already assuming that f is supported on
156
T. TAO
I with mean zero, it thus suffices to verify the estimate when f is the standard Haar wavelet φI : hφbI T ∗ (φI ), bχI i − hT ∗ (φbI φI ), bχI i = O(1). Throwing the T ∗ on the other side, we see the claim will follow from Cauchy-Schwarz and the estimates Z Z Z 2 2 |T (bχIlef t )| , |T (bχIright )| , |T (bχI )|2 . |I|. I
I
I
We just show the last estimate, as the first two are proven similarly. From weak boundedness we have Z 0 T (bχI )b . |I|1/2 . I
0
Since b is in L estimate
∞
0
with |[b ]I | ∼ 1, it thus suffices to prove the BMO Z |T (bχI ) − [T (bχI )]I |2 . |I|. I
On the other hand, since T (b) is in BMO by hypothesis, we have Z |T (b) − [T (b)]I |2 . |I|. I
The claim follows since T (b(1 − χI )) is constant on I by (4). From Lemma 3.4 one can show (as in the proof of Theorem 1.2) that the wavelet coefficients Wb (bT ∗ f )(I) obeys the Carleson measure estimate X |Wb (bT ∗ f )(J)|2 . |I|. J⊆I
(As in the proof of Theorem 1.2, it will be convenient to first remove the intervals where the mean of T (bχI ) is large, in order to estimate T (bχJ ) by T (bχI ) effectively.) From this Carleson measure estimate, (15) and the mean-zero property of the wavelets ψJb = bφbJ , we can show that T ∗ f ∈ BM O. Thus T ∗ maps L∞ to BMO. Similar arguments give the bound for T , and the Lp bounds then follow from interpolation as before. This allows one to prove the global T (b) theorem, Theorem 1.4. We now briefly indicate how the above proof of the global T (b) theorem can be localized to yield Theorem 1.5; the details though are rather complicated and can be found in [4]. In the above argument we gave pointwise estimates on wavelet coefficients Wb (bT ∗ f )(J), which were in turn used to prove Carleson measure estimates. Now, however, we do not have a single b to work with, instead having a localized bI assigned
THE T (b) THEOREM AND ITS VARIANTS
157
to each interval I. We would now like to have a Carleson measure estimate of the form X (16) |WbI (bI T ∗ f )(J)|2 . |I|. J⊆I
Suppose for the moment that we could attain (16). If bI obeyed the pseudo-accretivityR condition |[bI ]J | & 1 for all J ⊆ I, and the boundedness condition J |bI |2 . |J| for all J ⊆ I, then one could use (15) to then show the BMO estimate Z |T ∗ f − [T ∗ f ]I |2 . |I| (17) I ∗
which would show that T mapped L∞ to BMO as desired. Unfortunately, our assumptions only give us that bI has large mean on all of I: |[bI ]I | & 1. This does not preclude the possibility of bI having small mean on some sub-interval J of I (for instance, bI could vanish on some sub-interval). R 2 However, because we have the L bound I |bI |2 . I, it does mean that [bI ]J cannot vanish for a large proportion of intervals J. To make this more rigorous, fix 0 < ε 1, and let ΩI ⊆ I be the union of all the intervals J ⊆ I for which |[bI ]J | . ε. Then by a simple covering argument (splitting ΩI as the disjoint union of the maximal dyadic intervals J in ΩI ) we have Z bI . ε|ΩI | . ε|I| ΩI
and thus (if ε is sufficiently small) Z bI & |I|. I\ΩI
From Cauchy-Schwarz and the L2 bound on bI we thus have |I\ΩI | & |I|. Thus there is a significant percentage of I for which one does not have the problem of bI having small mean. Furthermore, by a similar argument one can also show that on a slightly smaller perR significant 2 centage of I, one also does not have the problem of J |bI | being much larger than J. Thus there is a non-zero “good” portion of I where the function bI behaves as if itwere pseudo-accretive, and on which Carleson estimates such as (16) would yield BMO bounds. The proof of (17) then proceeds by first estimating the integral on this “good” portion of I, and using a standard iteration argument to deal with the remaining
158
T. TAO
“bad” portion of I, which has measure at most c|I| for some c < 1; the point being that the geometric series 1 + c + c2 + . . . converges. Details can be found in [4]; similar arguments can also be found in [1], [6] (and indeed a vector-valued, higher-dimensional version of this trick of removing the small-mean intervals is crucial to the resolution of the Kato square root problem in higher dimensions [1], [2]). It remains to prove (16), at least when J is restricted to the set where bI behaves like a pseudo-accretive function. Obviously one would like to have Lemma 3.4 holding (but with b, b0 replaced by bI , b0I of course). However, an inspection that to do this one needs R of 2the proofR shows 0 2 bounds of the form J |bI | . |J|, J |bI | . |J|, and |[b0I ]J | & 1; in other words bI and b0I have to behave like pseudo-accretive functions2 on J. Fortunately, by arguments similar to the ones given above, one can show that the bad intervals J, for which the above bounds fail, do not cover all of I, and there is a fixed proportion of I which is “good”, in the sense that (16) holds when localized to this good set. One then performs yet another iteration argument on the “bad” portion of I to conclude (16). This concludes the sketch of the proof of Theorem 1.5 in the case of perfect dyadic Calder´on-Zygmund operators; further details can be found in [4]. It is an interesting question as to whether this proof technique can also produce localized estimates, perhaps similar to Theorem 1.3. One model instance of such a theorem is Theorem 29 of [9], which says informally that if E is a compact subset of an Ahlfors-David regular set with positive analytic capacity, then E has large intersection with another Ahlfors-David regular set for which the Cauchy integral is bounded. (This is somewhat analogous with a hypothesis that T (b), T ∗ (b) ∈ BM O for some b which is sometimes, but not always, pseudo-accretive, and a conclusion that T is bounded when restricted to a reasonably large set Ω). 4. Acknowledgements This expository paper is based on joint work [4] with Pascal Auscher, Steve Hofmann, Camil Muscalu, and Christoph Thiele. The author also thanks Mike Christ for some very helpful suggestions. The author is a Clay Prize Fellow and is supported by the Packard Foundation. 2Actually,
one also needs lower bounds for the magnitude of of [b0I ]Jlef t and This introduces some technicalities involving “buffer” intervals - intervals J for which bI has large mean, but such that bI has small mean on one of the subintervals Jlef t , Jright . These intervals are rather annoying to deal with but are fortunately not too numerous; see [4].
[b0I ]Jright .
THE T (b) THEOREM AND ITS VARIANTS
159
References [1] Auscher, P., Hofmann, S., McIntosh. A., Tchamitchian, P., The Kato square root problem for higher order elliptic operators and systems on Rn , submitted. [2] Auscher, P., Hofmann, S., Lacey, M. McIntosh. A., Tchamitchian, P., The solution of the Kato square root problem for second order elliptic operators on Rn , submitted. [3] Auscher, P., Hofmann, S., Lewis, J.. A., Tchamitchian, P., Extrapolation of Carleson measures and the analyticity of Kato’s square root operators, to appear, Acta Math. [4] Auscher, P., Hofmann, S., Muscalu, C., Tao, T., Thiele, C., Carleson measures, trees, extrapolation, and T (b) theorems, submitted. [5] Auscher, P, Tchamitchian, P., Bases d’ondelettes sur les courbes corde-arc, noyau de Cauchy et espaces de Hardy associ´es, Rev. Mat. Iberoamericana 5 [1989], 139–170. [6] Auscher, P, Tchamitchian, P., Square root problem for divergence operators and related topics. Ast´erisque No. 249, (1998), viii+172 pp. [7] Calder´ on, C., On commutators of singular integrals, Studia Math. 53 [1975], 139–174. [8] Carleson, L. On convergence and growth of partial Fourier series, Acta. Math. 116 135–157, [1966]. [9] Christ, M., A T (b) theorem with remarks on analytic capacity and the Cauchy integral. Colloq. Math. 60/61 [1990], 601–628. [10] Coifman, R.R., Jones, P., and Semmes, S. Two elementary proofs of the L2 boundedness of Cauchy integrals on Lipschitz curves. J. Amer. Math. Soc. 2 553–564 [1989]. [11] David, G., Wavelets and singular integrals on curves and surfaces. , Lecture Notes in Mathematics, 1465. Springer-Verlag, Berlin, 1991. [12] David, G., Journ´e, J., A boundedness criterion for generalized Calder´ on- Zygmund operators, Ann. of Math. 120 [1984], 371–397. [13] David, G., Journ´e, J., and Semmes, S., Op´erateurs de Calder´ on-Zygmund, fonctions para-accr´etives et interpolation, Rev. Mat. Iberoamericana 1 [1985], 1–56. [14] David, G. and Semmes, S., Analysis of and on uniformly rectifiable sets, Mathematical Surveys and Monographs, 38. American Mathematical Society, Providence, RI, 1993. xii+356 pp. ISBN: 0-8218-1537-7 [15] Fefferman, C. Pointwise convergence of Fourier series, Ann. of Math. (2) 98, pp. 551–571 [1973]. [16] C. Fefferman, Inequalities for strongly singular convolution operators, Acta Math. 124 (1970), 9–36. [17] Fefferman, C. and Stein, E. M., H p spaces of several variables, Acta Math. 129 [1972], 137–193. [18] Garnett, J., Jones, P., BMO from dyadic BMO, Pacific J. Math. 99 [1982], 351–371. [19] Lacey, M. and Thiele, C. Lp estimates on the bilinear Hilbert transform for 2 < p < ∞. Ann. Math. 146, pp. 693-724, [1997] [20] Lacey, M. and Thiele, C. On Calderon’s conjecture. Ann. Math. 149, pp. 475496, [1999]
160
T. TAO
[21] Lacey, M. and Thiele, C. A proof of boundedness of the Carleson operator, Math. Res. Lett 7 361–370, [2000] [22] Muscalu, M., Tao, T., Thiele, C. Lp estimates for the “biest”, submitted, Math. Annalen. [23] Nazarov, F., Treil, S., Volberg, A., Accretive system T b theorems on nonhomogeneous spaces, preprint. [24] Nazarov, F., Treil, S., Volberg, A., T b theorems on nonhomogeneous spaces, preprint. [25] Semmes, S., Square function estimates and the T (b) theorem, Proc. Amer. Math. Soc. 110, 721–726 [1990] [26] Stein, E. Harmonic Analysis: Real Variable Methods, Orthogonality, and Oscillatory Integrals. Princeton University Press, Princeton, [1993] [27] Thiele, C. Time-Frequency Analysis in the Discrete Phase Plane, PhD Thesis, Yale 1995. [28] Thiele, C. On the Bilinear Hilbert transform. Universit¨at Kiel, Habilitationsschrift [1998] [29] Wolff, T.H. A note on interpolation spaces. Harmonic analysis (Minneapolis, Minn., 1981), pp. 199–204, Lecture Notes in Math., 908, Springer, Berlin-New York [1982] Department of Mathematics, UCLA, Los Angeles CA 90095-1555 USA E-mail address: [email protected]
SEPARATION OF GRADIENT YOUNG MEASURES AND THE BMO KEWEI ZHANG Dedicated to Professor Alan McIntosh for his 60th birthday Abstract. Let K = {A, B} ⊂ M N ×n with rank(A − B) > 1 and Ω ⊂ Rn be a bounded arcwise connected Lipschitz domain. We show that there is a direct estimate of the size of the -neighbor¯ (A) ∪ B ¯ (B) separates gradient hood K of K such that K = B 1,1 Young measures, that is, if (u ) ⊂ W (Ω, RN ) is bounded and j R dist(Du Ω R j , K )dx → 0 as j → ∞,R then up to a subsequence, ¯ (A))dx → 0 or ¯ (B))dx → 0. either Ω dist(Duj , B dist(Duj , B Ω
In this note I present some direct estimate on neighborhoods K of a two matrix set K = {A, B} that separate gradient Young measures. I also show how one can use the BM O seminorm of approximate solutions of linear elliptic systems to control the oscillation of sequences gradients approaching K . The note is based on a recent work [19] with a slightly different approach. In the case of the two matrix set, we can establish our main result (Theorem 1 below) without quoting a recent deep approximation theorem obtained by M¨ uller [14], instead, most of the tools we use will be standard in the calculus of variations. The problem we consider is motivated from the variational approach of material microstructure [4, 5], in particular, the study of metastablility, hysteresis and numerical analysis related to them. Let M N ×n be the space of N × n real matrices with N, n ≥ 2. Let Ω be a bounded arcwise connected Lipschitz domain throughout ∗ this note. We denote by * and →* weak convergence and weak-∗ convergence respectively. The characteristic function of a set V ⊂ Rn is denoted by χV . The following result is now well-known [4, 17]. Theorem A. Let A, B ∈ M N ×n with rank(A − B) > 1. Let (uj ) ⊂ W 1,p (Ω,RRN ) (1 ≤ p < ∞) be a bounded sequence such that limj→∞ Ω distp (Duj , {A, B}) → 0. Then, up to a subsequence either Duj → A a.e. or Duj → B a.e.. In the study of metastablility and hysteresis of material microstructure, Ball and James [6] established the following: 161
162
KEWEI ZHANG
Theorem B. If a compact set K = K1 ∪ K2 ⊂ M N ×n (K1 ∩ K2 = ∅) separates gradient Young measures in the sense that supp νx ⊂ K a.e. ⇒ supp νx ⊂ K1 or supp νx ⊂ K2 a.e. Then there exists > 0, such that K still separates gradient Young measures. In other words, K separates gradient Young measures if dist(Duj , K) → 0 in L1 (Ω) implies, up to a subsequence either dist(Duj , K1 ) → 0 in L1 (Ω) or dist(Duj , K2 ) → 0 in L1 (Ω). Theorem B then claims that there exists > 0 such that K still separates gradient Young measures. Theorem B was established by using a contradiction argument in [6]. In this note I give an estimate of > 0 for the special case K = {A, B} ¯ (A)∪ B ¯ (B) separates gradient with rank(A−B) > 1 such that K = B Young measures, i.e. we give an estimate of the size of the balls such that a sequence of gradients approaching two balls can only approach one. We have, Theorem 1. Let K = {A, B} ⊂ M N ×n with rank(A − B) > ¯ (A) ∪ B ¯ (B), and λmax be the largest eigenvalue of 1. Let K = B t (A − B) (A − B)/|A − B|2 . Then there exists an estimate of > 0 depending on n, N, |A − B|, λmax such that uj * u in W 1,1 (Ω, RN ), R dist(Duj , K )dx → 0 as j → ∞ implies that up to a subsequence, Ω either Z Z ¯ ¯ (B))dx = 0. lim dist(Duj , B (A))dx = 0, or lim dist(Duj , B j→∞
j→∞
Ω
Ω
In the language of gradient Young measures [11, 15], this means ¯ (A) a.e. or supp νx ⊂ K , a.e. x ∈ Ω implies either supp νx ⊂ B ¯ (B) a.e. supp νx ⊂ B We use the following standard tools to establish Theorem 1. (i) A non-negative quasiconvex function vanishing exactly on K [10, 17], (ii) homogeneous Young measures [15], (iii) BMO estimates for elliptic systems [9], (iv) Besicovitch’s covering lemma [8]. More precisely, we have an explicit non-negative quasiconvex function f satisfying 0 ≤ f (X) ≤ C(1 + |X|2 ),
f −1 (0) = K .
R So if Ω dist2 (Duj , K )dx → 0 and uj * u in W 1,2 , then by the wellknown weak lower semicontinuity theorem of Acerbi and Fusco [1], Z Z f (Duj )dx ≥ f (Du)dx. 0 = lim inf j→∞
Ω
Ω
SEPARATION OF GRADIENT YOUNG MEASURES
163
In the language of gradient Young measures, this gives Z Z Z Z f (λ)dνx (λ)dx = f (¯ νx )dx, where ν¯x = λdνx = Du(x), M N ×n
Ω
M N ×n
Ω
hence supp νx ⊂ K , and more importantly, we can locate the weak limit ν¯x = Du(x) ∈ K a.e. (1) Recall that a continuous function f : M N ×n → R is quasiconvex [13, 3] if Z f (A + Dφ) ≥ f (A)|Ω|, for A ∈ M N ×n , φ ∈ C01 (Ω, RN ). Ω p It is well R known that if 0 ≤ f (X) ≤ C(1+|X| ), the variational integral u → Ω f (Du)dx is sequentially weakly lower semicontinuous in W 1,p if and only if f is quasiconvex [1]. We may construct quasiconvex functions by calculating the quasiconvex envelope QF for a given function F : M N ×n → R: QF = sup{g ≤ F, g quasiconvex}. In our case, we let F (X) = dist2 (X, {A, B}), then QF can be explicitly calculated [10] and
QF (X) = F (X) = dist2 (X, {A, B}), if
dist2 (X, {A, B}) ≤ |A − B|2 (1 − λmax )/4.
Let f (X) = Q dist2 (X, {A, B}) − + ,
≤
|A − B|2 (1 − λmax ) , 4
then f ≥ 0 is quasiconvex and f −1 (0) = K . Next need the W 1,∞ -Gradient Young measure and the homogeneous Young measure [11, 15] to localize our problem: We only need a special case of the general theorem of gradient Young measures. Let K ⊂ M N ×n be compact, uj * u in W 1,1 (Ω, RN ) such that dist(Duj , K) → 0 in L1 . Then (I) up to a subsequence, there exists a family of probability measures (gradient Young measures) supp νx ⊂ K, a.e. x ∈ Ω, f (Duj ) * R f (λ)dνx weakly in L1 , for all continuous functions f satisfying K R |f (X)| ≤ C(|X|+1) and the weak limit can be identified as Du(x) = K λdνx := ν¯x a.e. (νx )x∈Ω is called the family of W 1,∞ -gradient Young measures generated by (a subsequence) of gradients (Duj ). (II) There is a bounded sequence vj ∈ W 1,∞ such that (Dvj ) generates the same family of Young measures (νx ), kDvj − Duj kL1 → 0, R ∗ f (Dvj ) →* K f (λ)dνx for all continuous functions f .
164
KEWEI ZHANG
(III) For a.e. x0 ∈ Ω, there exists a bounded sequence (φk ) in W01,∞ (D, RN ) where D is the unit disk in Rn , such that the corresponding gradient Young measures {ˆ νy } of the sequence (Du(x0 ) + Dφk ) satisfy νˆy = νx0 a.e. y ∈ D. We call νˆy the homogeneous Young measure and simply denote it by ν := νˆy , y ∈ D. The homogeneous Young measure enables us to localize our problem. We may decompose Theorem 1 into Lemma 1. (‘Existence’) Suppose ν is a homogeneous Young measure ¯ (A). Then supp ν ⊂ B ¯ (A). satisfying supp ν ⊂ K and ν¯ = X ∈ B ¯ (A) implies Lemma 1 implies that for a fixed x, supp νx ⊂ K , ν¯x ∈ B ¯ supp νx ⊂ B (A)). Lemma 2. (‘Regularity’) There is an estimate of > 0 depending on |A−B|, λmax , n and N such that if Du(x) ∈ K a.e. in Ω. Then either ¯ (A) a.e. or Du(x) ∈ B ¯ (B) a.e. Du(x) ∈ B ¯ (A) a.e. or ν¯x ∈ B ¯ (B). Lemma 2 shows that if ν¯x ∈ K then ν¯x ∈ B Combining (1), Lemma 1 and Lemma 2, we reach the conclusion of Theorem 1. To prove Lemma 1 we need to control the large scale oscillation of the gradients of a sequence of mappings with a fixed affine boundary condition. For Lemma 2, we need to rule out large scale oscillation of the gradient of a fixed mapping without prescribed boundary condition. Note that in either cases the best possible we can reach is some partial rigidity of the gradients. The tool we use is the Schauder estimates in BM O and Campanato spaces for linear elliptic systems with constant coefficients. Notice the ellipticity property of linear subspaces of M N ×n without rank-one matrices [2]. Let E = span[A−B]. Then E is a subspace without rank-one matrices. Let E ⊥ be the orthogonal complement of E and PE ⊥ be the orthogonal projection to E ⊥ , then there is some constant c0 > 0 such that for any rank-one matrix a ⊗ b, |PE ⊥ (a ⊗ b)|2 ≥ c0 |a|2 |b|2 , where a ∈ RN and b ∈ Rn . Let us recall some basic facts about elliptic system with constant coefficients [9]: ∞ i j i div Aij αβ Dβ u = div Fα in Ω with Fα ∈ L .
(2)
We say that the system satisfy the Legendre-Hadamard strong ellipticity condition if i j 2 2 n N Aij αβ ξα ξβ η η ≥ λ0 |ξ| |η| , ξ ∈ R , η ∈ R .
(3)
i j 2 2 We denote by Λ0 > 0 a constant such that Aij αβ ξα ξβ η η ≤ Λ0 |ξ| |η| .
SEPARATION OF GRADIENT YOUNG MEASURES
165
Example. Let E ⊂ M N ×n be a subspace without rank-one matrices. Then the second order elliptic system div PE ⊥ Du = div F satisfies (3). In our case λ0 = 1 − λmax , and Λ0 = 1. Let u ∈ W 1,2 be a weak solution of (2). The following are some estimates of Du. (A) Interior Estimate ¯R (x0 ) ⊂ Let x0 ∈ Ω and 0 < ρ < R such that Bρ (x0 ) ⊂ BR (x0 ) ⊂ B Ω, Z −
|Du − [Du]x0 ,ρ |2 dx
Bρ (x0 )
Z ρ τ 2 2 − |Du − [Du]x0 ,R | dx + [F ]L2,n (Ω) , ≤C R BR (x0 )
(4)
where C = C(n, N, λ0 , Λ0 ) > 0, τ = τ (n, N, λ0 , Λ0 ), 0 < τ < 2, and [F ]L2,n (Ω) is the Campanato seminorm. (B) Global Estimate Under Dirichlet condition u|∂Ω = 0, kDukBM O(Ω) ≤ CkF kL∞ .
(5)
We prove Lemma 2 first which depends on the interior estimate (a) above. Proof of Lemma 2. Without loss of generality, we may assume that A = 0, E = span[B]. This can be done by a simple translation in M N ×n . Note that |PE ⊥ (Du)| ≤ 2, hence u ∈ W 1,∞ (Ω, RN ) is a weak solution of where F = PE ⊥ (Du), kF kL∞ ≤ 2. ¯ (B)}. Du(x) ∈ K implies that Let V = {x ∈ Ω, Du(x) ∈ B div PE ⊥ (Du) = div F,
Du = BχV + H,
kHkL∞ ≤ 4.
(6)
Then (a) implies, for a fixed x0 ∈ Ω and 0 < ρ < R such that B2R (x0 ) ⊂ Ω, Z − |Du−[Du]x0 ,ρ |2 dx Bρ (x0 ) Z ρ τ 2 2 − |Du − [Du]x0 ,R | dx + [F ]L2,n (Ω) ≤C R BR (x0 ) ρ τ ≤C (|B| + 2)2 + C2 ≤ C2 R
166
KEWEI ZHANG
if ρ/R is sufficiently small. Consequently, for every x0 ∈ Ω, there is a small cubes Q(x0 , r) ⊂ Ω centered at x0 with side-length r > 0 depending on x0 and Ω, kDuk2BM O(Q(x0 ,r) ≤ C2 .
(7)
By applying the definition of BM O and (3), we have, for each cube Q ⊂ Q(x0 , r) and let G(Q) =
|Q ∩ V | , |Q|
then
G(Q)(1 − G(Q)) ≤
C 2 3 < , 2 |B| 16
as long as > 0 is small. Hence for Q ⊂ Q(x0 , r), 1 3 or G(Q) ≥ . (8) 4 4 Now we use the Intermediate Value Theorem and a density argument to finish the proof. Without loss of generality, we assume x0 ∈ Ω is a point of density 1 for V . Then we show that there is no points of density 1 for Ω \ V in Ω. Since x0 is a point of density 1 for V , there is a cube Q0 ⊂ Q(x0 , r) centered at x0 , such that G(Q0 ) > 3/4. We first show that there is no point of density 1 for Q0 \V in Q0 . Otherwise, let x1 ∈ Q0 be an interior point such that there is some Q1 ⊂ Q0 centered at x1 and satisfies G(Q1 ) < 1/4. Then we may construct a continuous family of decreasing cubes Q(t) in Q0 such that Q(0) = Q0 and Q(1) = Q1 . It is then easy to see that t → G(Q(t)) is a continuous function. By the Intermediate Value Theorem, there is some cube Q(t0 ) ⊂ Q0 ⊂ Q(x0 , r), such that G(Q(t0 )) = 1/2. This contradicts to (8). Since Ω is arcwise connected, for each x ∈ Ω, there is a piecewise affine curve γ : [0, 1] → Ω, such that dist(γ, ∂Ω) = δ0 > 0, γ(0) = x0 , γ(1) = x. If we choose r > 0 sufficiently small and let Q(t) be the cube centered at γ(t) with radius 0 < r < δ0 , we may claim that (8) holds for Q ⊂ Q(t), 0 ≤ t ≤ 1. Then it is easy to see that x is not a point of density 1 for Ω \ V , hence Du(x) ∈ B (B) a.e. either
G(Q) ≤
Proof of Lemma 1. We may assume A = 0 as before and supp ν ⊂ K ¯ (0). Let X +Duj generates ν with uj ∈ W01,∞ bounded. and ν¯ = X ∈ B Define PE ⊥ (Duj ), if |PE ⊥ (Duj )| ≤ 4, Fj = 0, otherwise, where E = span[B]. Solving div PE ⊥ (Dvj ) = div Fj in D, vj |∂D = 0, kDvj kBM O(D) ≤ C,
kDuj − Dvj kL2 → 0
(9)
SEPARATION OF GRADIENT YOUNG MEASURES
167
Let Wj = {x ∈ D, dist(Dvj , K) ≥ 4} so |Wj | → 0 and let Vj = {x ∈ D, |Dvj − B| < 4}, then Dvj = BχVj + Dvj χWj + O(). Let Gj (Q) = |Vj ∩ Q|/|Q|, then (9) implies Z 2 Gj (Q)(1 − Gj (Q)) ≤ C + C − (1 + |Dvj |2 )χWj dx. (10) Q
R
Also notice that D Dvj dx = 0 which implies Gj (D) ≤ C < 1/4 for large j. ¯ ) → 0 in Our aim is to show that |Vj | → 0 so that dist2 (X + Dvj , B 1 ¯ (0). Now we use a slightly different argument as in L , hence ν supp B R the proof of Lemma 1 by using Wj (1 + |Dvj |2 )dy to bound |Vj |. For each point x ∈ D of density 1 for Vj , there exists a cube Q ⊂ D centered at x such that Gj (Q) > 3/4. Note that Gj (D) < 1/4, we can then prove that there is an open cube Qx containing Q in D such that Gj (Qx ) = 1/2 which maximize the left hand side of (10). If we further require that 1/4 − C2 := γ > 0, then from (10), Z 1 2 γ|Qx | = |Qx | − C ≤ C (1 + |Dvj |2 )χWj dy. (11) 4 Qx Clearly {Qx } is a covering of the points of density 1 for Vj by open cubes. By Besicovitch’s covering lemma (see e.g. [8]), it is then easy to prove that Z (1 + |Dvj |2 )dy → 0.
γ|Vj | ≤ C Wj
¯ (0). Therefore |Vj | → 0 so that supp ν ⊂ B
Theorem 1 can be generalized to any finite sets contained in a subspace without rank-one matrices [19]. For a finite set K = {Ai } ⊂ M N ×n , we define the diameter of K dK = max{|Ai − Aj |, i 6= j}, and the smallest distance gK = min{|Ai − Aj |, i 6= j}. Theorem 2. Suppose E ⊂ M N ×n be a linear subspace without rankone matrices. Let K = {Ai } ⊂ E be a finite subset. Let λE = min{|PE ⊥ (a ⊗ b)|2 , a ∈ RN , b ∈ Rn , |a| = |b| = 1}, |PE (a ⊗ b)|2 1 − λE 1 = inf = . 2 µE |a|=|b|=1 |PE ⊥ (a ⊗ b)| λE Then there exists an estimate of > R0 depending on dK , gK , λE , µE , n and N , such that uj * u in W 1,1 , Ω dist(Duj , K ) → 0 implies, up to a subsequence, for some fixed Ai0 ∈ K, Z ¯ (Ai0 )) → 0. lim dist(Duj , B j→∞
Ω
168
KEWEI ZHANG
Remark 1. Theorem 2 was proved in [19] by using S. M¨ uller’s improved approximation lemma for sequences of gradients approximating a compact set [14]. In the special case of Theorem 1, we may avoid using this result, instead, establishing Lemma 1 directly form the global estimate for the Dirichlet problem. SmRemark 2. For the incompatible multi-elastic well structure K = i=1 SO(2)Hi with Hi positive definite, the theory for linear elliptic system still works provided that the wells are sufficiently ‘flat’ [20]. Remark 3. For the two well structure K = SO(n) ∪ SO(n)H, it is known [12, 16] that under a technical assumption on H, the compactness result holds, that is, if dist2 (Duj , K) → 0 in L1 , then there exists some A ∈ K, such that Duj → A a.e. A nonlinear elliptic system is involved. However, I do not know any interior BM O estimates for the elliptic system div A(Du) = div f , u|∂Ω = 0, where c|Y |2 ≤ DA(X)Y Y ≤ C|Y |2 , kf kL∞ ≤ ? As remarked in [12], if H = λI with λ > 0, λ 6= 1, and I being the identity matrix, one may simply use the n-Laplace operator to study convergent sequences of gradients to K, that is, div |Du|n−2 Du = div F. However, even for this explicit system, I do not know any BM O estimate of the weak solutions in W 1,n given that kF kL∞ is small. References [1] E. Acerbi and N. Fusco. Semicontinuity problems in the calculus of variations, Arch. Rational Mech. Anal. 86, 125–145, 1984. [2] K. Bhattacharya and N. B. Firoozye, R. D. James, R. V. Kohn. Restrictions on Microstructures, Proc. Royal Soc. Edinburgh A 124, 843–878, 1994. [3] J. M. Ball. Convexity conditions and existence theorems in nonlinear elasticity, Arch. Rational Mech. Anal. 63, 337–403, 1977. [4] J. M. Ball and R. D. James. Fine phase mixtures as minimizers of energy, Arch. Rational Mech. Anal. 100, 13–52, 1987. [5] J. M. Ball and R. D. James. Proposed experimental tests of a theory of fine microstructures and the two-well problem, Phil. Royal Soc. Lon. 338A 389– 450, 1992. [6] J. M. Ball and R. D. James. Personal communication, 1993. [7] B. Dacorogna. Direct Methods in the Calculus of Variations, Springer-Verlag, 1989. [8] L. C. Evans and R. F. Gariepy. Measure Theory and Fine Properties of Functions, CRC Press, 1992. [9] M. Giaquinta, Introduction to regularity theory for nonlinear elliptic systems, Birkh¨ auser Verlag, Basel, 1993. [10] R. V. Kohn. The relaxation of a double-well energy, Cont. Mech. Therm. 3, 981–1000, 1991.
SEPARATION OF GRADIENT YOUNG MEASURES
169
[11] D. Kinderlehrer and P. Pedregal. Characterizations of Young measures generated by gradients, Arch. Rational Mech. Anal 115 329–365, 1991. [12] J. P. Matos. Young measures and the absence of fine microstructures in a class of phase transitions, Eur. J. Appl. Math. 3, 31–54, 1992. [13] C. B. Jr Morrey. Multiple integrals in the calculus of variations, Springer, 1966. [14] S. M¨ uller. A sharp version of Zhang’s theorem on truncating sequences of gradients, Trans. AMS 351, 4585–4597, 1999. [15] P. Pedregal. Parametrized measures and variational principles. Progress in Nonlinear Differential Equations and their Applications. 30 Basel: Birkh¨auser, 1997. ˇ ak. On the problem of two wells. D. Kinderlehrer (ed.) et al., Mi[16] V. Sver´ crostructure and phase transition, New York, NY: Springer-Verlag. IMA Vol. Math. Appl. 54, 183–189, 1993. [17] K.-W. Zhang. A construction of quasiconvex functions with linear growth at infinity, Ann. Sc. Norm. Sup. Pisa Serie IV XIX, 313–326, 1992. [18] K.-W. Zhang. A two-well structure and intrinsic mountain pass points, Calc. of Var. PDEs 13, 231–264, 2001. [19] K.-W. Zhang. On separation of gradient Young measures. To appear in Calc. of Var. PDEs [20] K.-W. Zhang, paper in preparation. School of Mathematical Sciences, University of Sussex, Falmer, Brighton, BN1 9QH E-mail address: k.zhang sussex.ac.uk